nr-bundles-public

A curated, multi-modal dataset of blockchain validator infrastructure under attack and benign workloads. 918 bundles across 37 chains and network stacks (Aptos, Bitcoin, bnb-smart-chain, Cardano, casper, celestia, conflux, Cosmos, Dogecoin, Ethereum, ethereum-consensus-layer, Filecoin, http2, http3, ic, icon, IOTA, ipfs, kaspa, libp2p, Litecoin, Monero, Near, optimism, polkadot, polkadot-substrate, polygon-pos, qtum, quic, Solana, Solana (Agave), sonic-fantom, starknet, Sui, walrus, XRP, zcash), all thirteen families in the taxonomy (twelve attack classes plus benign), and 200 attack primitive_ids — including faithful wire-level reproductions of 148 publicly disclosed CVEs and security advisories in blockchain node software. Released as the public sample of NullRabbit's bundle corpus, which underpins NullRabbit Labs' ML research on autonomous defence for decentralised networks.

What this dataset is

Most security datasets on the Hub are either application-layer (smart contract vulnerabilities, exploit code snippets) or generic (enterprise network traffic, malware signatures). The infrastructure layer where blockchain validators actually run (the RPC surface, the gossip layer, the consensus daemon, the indexer) has no shared, ML-trainable representation. This dataset is the first public sample of one.

Each entry is a bundle: a multi-modal observation of validator infrastructure under a single workload. Bundles capture network packets, host telemetry, application metrics, and observed responses simultaneously, under a chain-agnostic schema. The format is open and specified at nr-bundle-spec (MIT-licensed). Anyone can produce v1 bundles for any chain.

This release publishes bundles drawn from the NullRabbit corpus, with archive-provenance recorded per-bundle (see the composition breakdown above). Two provenance classes are represented: original (NullRabbit's own disclosed findings, each with a published operator advisory) and public-cve-replication (faithful reproductions of already-public CVEs / security advisories, each linking its original disclosure via provenance.public_source). The public-cve-replication bundles are the browsable, per-bundle form of data that already underpins the public nr-network-known-class-detector model — this release publishes them as inspectable bundle data, not as a first appearance on the Hub. The bundles selected demonstrate format coverage across the populated vulnerability families and provide attack/benign training material for cross-chain anomaly detection research.

Bundle structure

Each bundle is a directory containing a manifest plus up to six modality files. Per-bundle modality presence is reflected in the manifest's files block; bundles do not contain modalities that were not captured during the workload, and the manifest is the ground-truth for which files are present.

File	Modality	Present in this release
manifest.json	Schema-pinned metadata + provenance hashes	All bundles
packets.pcap	Raw network-layer packet capture	Dropped from every bundle (post-term cleartext concern; see Limitations)
pcap_pre_termination.pcap	Pre-termination TLS frames (encrypted; IP/TCP headers visible)	Bundles at fidelity_class: lab-tls-fronted only (17 bundles in this release)
host.parquet	Host telemetry (CPU, memory, IO, syscall counters)	All bundles
app.parquet	Application-level metrics from the validator daemon	Bundles where the validator exposed a metrics endpoint (most bundles; empty placeholders elsewhere)
protocol.parquet	Chain-agnostic consensus/protocol signal time-series	Bundles where consensus/protocol signals were captured (manifest protocol_parquet=True; empty placeholders for some passive-workload benigns)
responses.parquet	RPC/API response metadata (no bodies)	Bundles where the validator served RPC traffic during the capture window (manifest responses_parquet=True); absent or zero-rows for passive-workload benigns (sui_BENIGN_passive_fullnode, solana_BENIGN_validator_passive)
vectors.parquet	Computed feature vectors derived from the above	Not populated in this release. Reserved schema slot; consumers should derive feature vectors at inference time using the bundle-spec reference parser plus their own feature pipeline

The manifest pins primitive_id, family_id, traffic_source, fidelity_class, target_authorisation, and provenance hashes for every file in the bundle. The schema is chain-agnostic: no field assumes a specific blockchain. The BundleFiles block (packets_pcap, host_parquet, app_parquet, protocol_parquet, responses_parquet, vectors_parquet) is the authoritative present/absent contract; readers should check the manifest before reading each modality file.

Bundles in this release carry one of three fidelity_class values: lab (controlled reproducer environment, no TLS termination), lab-tls-fronted (controlled reproducer with an nginx TLS-termination front; pre-term TLS pcap is retained, post-term cleartext pcap is dropped), and proxy (the workload is exercised through a protocol proxy / synthetic-client harness standing in for the daemon surface; no TLS front, raw pcap dropped). All bundles have target_authorisation: self-owned (every validator targeted was operated by NullRabbit).

Full schema: nr-bundle-spec v0.2.0, which defines bundle v1 format.

Family taxonomy

Bundles are classified by family, defined by attack mechanism rather than by chain. Because the format and the mechanism are shared across chains, cross-chain generalisation becomes testable: you can hold out an entire chain and measure whether a model trained on the others recovers a family. On this corpus it does not transfer yet: held-out-chain family macro-F1 is 0.17 (Sui held out) / 0.35 (Solana held out) — measured on the original eight-chain cut, before the public-CVE-replication additions — against a 7-class random floor of ~0.14. Cross-chain transfer is the open question the shared format lets you pose and measure, not a result this dataset demonstrates; closing the gap needs more chains per family, not more features.

These are validator / node-software vulnerability classes only. On-chain contract / DeFi-economic findings are a separate, out-of-scope namespace and are never used here. The full taxonomy defines thirteen families (twelve attack classes plus benign). All thirteen are populated in this release:

Family	Mechanism
response_amp	Asymmetric output:input byte ratio
compute_amp	Asymmetric work:request CPU ratio
rate_limiter_bypass	Behaviour conformant with rate limits but resource-exhausting
reconnaissance	Information gathering for follow-on attacks
service_misconfig	Exploitable default or operator configurations
auth_bypass	Authentication/authorisation circumvention
connection_exhaustion	Per-connection/subscription resource leak, not released on disconnect
memory_amp	Unbounded accumulation of process memory under attacker input
consensus_abuse	Consensus-reactor message flood (valid-shape, unsigned) driving per-message CPU before sig-verify
subscription_cpu_amp	Wide/streaming subscription with disproportionate per-event server CPU
state_import_abuse	Malformed/oversized state artefact crashes the snapshot/bootstrap import path
gossip_abuse	Unauthenticated peer-to-peer / gossip surface floodable without admission control or rate-limits
benign	Validator under normal workload (negative class)

What's included

918 bundles total: 911 attack, 7 benign. By provenance class: 328 original (NullRabbit's own disclosed findings, each with a published operator advisory), 23 legacy-sample (the 2026-05-13 Solana/Sui curation), and 567 public-cve-replication (faithful reproductions of already-public CVEs / advisories). The raw packets.pcap is dropped from every bundle (BundleFiles.packets_pcap=False); no pre-termination pcap variant is retained.

The Pcap modality column in the tables below specifies which network-layer file is present on disk per bundle: pre-term TLS means pcap_pre_termination.pcap is retained (encrypted TLS frames; IP/TCP headers visible); none means the raw packets.pcap was dropped and no pre-term variant exists (BundleFiles.packets_pcap=False in the manifest).

Disclosed findings and legacy sample (137 bundles)

These are NullRabbit's own original findings — each with a published operator advisory — plus the legacy 2026-05-13 Solana/Sui sample.

Solana (23 bundles)

Primitive	Family	Bundles	Fidelity	Pcap modality	Coverage
SOL_F10_multi_get_accounts_amp	response_amp	4	lab-tls-fronted	pre-term TLS	4 postures
sol_snapshot_oversized_datalen_indexgen_panic	state_import_abuse	1	lab	none	attack; oversized AppendVec data_len → index-gen panic on snapshot import
SOL_F14_simulate_transaction_sync_wedge	compute_amp	4	lab-tls-fronted	pre-term TLS	4 postures
SOL_P07_get_program_accounts_filter_miss	compute_amp	5	lab	none	5 postures
SOL_RC_nmap_slow	reconnaissance	1	lab	none	saturating posture
SOL_MC_grafana_anon	service_misconfig	1	lab	none	saturating posture
SOL_MC_ssh_pwauth	service_misconfig	1	lab	none	saturating posture
SOL_H01_admin_rpc_probe	auth_bypass	1	lab	none	saturating posture
SOL_GR01_simulate_compute_flood	rate_limiter_bypass	1	lab	none	saturating posture
solana_BENIGN_organic_rpc	benign	2	lab-tls-fronted	pre-term TLS	mixed postures
solana_BENIGN_validator_passive	benign	2	lab	none	mixed postures

The SOL_F10 / SOL_F14 / SOL_P07 primitives correspond to the findings published as NR-2026-001 (Agave RPC architectural findings, 2026-05-12).

The two Solana benign primitives have different fidelity classes because they observe different workloads through different capture paths. solana_BENIGN_organic_rpc is RPC-client traffic against the validator (JSON-RPC over TCP); the Step-11 capture sweep included a TLS-fronted variant where nginx terminates TLS in front of the validator, enabling the pre-term TLS pcap modality. solana_BENIGN_validator_passive is passive consensus-layer observation (validator running idle, no RPC clients); no TLS-fronting capture path was run for this workload, so the lab variant is the only one available.

Sui (14 bundles)

Primitive	Family	Bundles	Fidelity	Pcap modality	Coverage
sui_F10_multi_get_objects_amp	response_amp	3	lab-tls-fronted	pre-term TLS	3 postures
sui_F14_devinspect_tokio_wedge	compute_amp	3	lab-tls-fronted	pre-term TLS	3 postures
sui_h02_state_sync_flood	gossip_abuse	3	proxy	none	attack; anemo state_sync P2P admits unauthenticated peers (no RequireAuthorizationLayer, rate-limits off) → floodable (SUI_H02; NR-2026-011); 3 postures
sui_f10_multiget_response_amp	response_amp	1	lab	none	attack; current-namespace capture of the F10 multiGetObjects egress amp (cf. legacy sui_F10_multi_get_objects_amp above)
sui_subscription_permit_leak	connection_exhaustion	1	lab	none	attack; subscription permit + streamer-map leak
sui_BENIGN_passive_fullnode	benign	1	lab	none	passive workload
sui_BENIGN_reproducer_pipeline	benign	1	lab-tls-fronted	pre-term TLS	reproducer baseline
sui_BENIGN_validator_under_load	benign	1	lab	none	active workload

IOTA (22 bundles)

Primitive	Family	Bundles	Fidelity	Pcap modality	Coverage
iota_grpc_stream_cap_dos	connection_exhaustion	16	lab	none	attack; StreamCheckpoints subscription-semaphore (1024) hold-shape exhaustion, saturating posture with a 640–1024 stream-count sweep
iota_f10_multiget_amp	response_amp	4	lab	none	attack; iota_multiGetObjects showBcs full-BCS egress amplification (IOTA_F10; NR-2026-010); 4 postures
iota_grpc_h2_multiplex_oom	memory_amp	1	lab	none	attack; unbounded HTTP/2 concurrent-stream OOM on the public gRPC surface
iota_s1_widefilter_subscribe_cpu	subscription_cpu_amp	1	lab	none	attack; wide event-filter subscription → exponential per-event CPU

The three 2026-07-02 additions (sui_f10_multiget_response_amp, sui_subscription_permit_leak, iota_grpc_h2_multiplex_oom) are NullRabbit original-research findings (source_class: original, ground_truth_label: attack), each also the subject of a published operator advisory (NR-2026-004, NR-2026-005, NR-2026-003 respectively). They add IOTA as a third chain and populate the connection_exhaustion and memory_amp families.

The two 2026-07-03 additions (sol_snapshot_oversized_datalen_indexgen_panic, iota_s1_widefilter_subscribe_cpu) populate the two families promoted into the taxonomy in nr-bundle-spec v0.2.0: state_import_abuse (a malformed snapshot artefact that crashes the bootstrap import path) and subscription_cpu_amp (a wide event-filter subscription driving exponential per-event CPU).

Cosmos (45 bundles)

Primitive	Family	Bundles	Fidelity	Pcap modality	Coverage
cometbft_proposal_flood	consensus_abuse	4	lab	none	4 postures; ProposalMessage flood (DataChannel 0x21)
cometbft_vote_flood	consensus_abuse	4	lab	none	4 postures; VoteMessage flood (VoteChannel 0x22)
cometbft_blocksync_flood	memory_amp	4	lab	none	4 postures; oversized BlockResponse accumulation (BlockSync 0x40)
cometbft_mconn_handshake_burn	compute_amp	17	lab	none	attack; SecretConnection STS handshake pre-auth X25519+Ed25519 crypto burn on :26656; 4 postures (saturating, distributed, low-volume, mimicry)
cosmos_grpc_stream_flood	connection_exhaustion	16	lab	none	attack; cosmos-sdk gRPC server built without MaxConcurrentStreams → unbounded held-stream per-stream pin on :9090; saturating posture with a 512–2048 stream-count sweep (COSMOS_SDK_GRPC_STREAM_FLOOD / C11; NR-2026-018)

The three CometBFT consensus-channel floods are one finding family (operator advisory NR-2026-007): valid-shape unsigned consensus messages queued before signature verification. source_class: original; MEDIUM — a multi-validator quorum keeps producing blocks under the attack. Mechanistically they split across two families as the manifests record: cometbft_proposal_flood and cometbft_vote_flood drive single-node CPU (consensus_abuse), while cometbft_blocksync_flood accumulates process memory via oversized BlockResponses (memory_amp).

The 2026-07-04 additions (cometbft_mconn_handshake_burn, iota_grpc_stream_cap_dos) are NullRabbit original-research findings (source_class: original, ground_truth_label: attack). cometbft_mconn_handshake_burn is a pre-authentication SecretConnection STS-handshake crypto burn on the CometBFT P2P port (COSMOS_MCONN_PREAUTH); iota_grpc_stream_cap_dos is a hold-shape exhaustion of the IOTA gRPC StreamCheckpoints subscription semaphore (IOTA_GRPC_STREAM_CAP_DOS, HIGH). They deepen the compute_amp and connection_exhaustion families respectively.

Ethereum (17 bundles)

Primitive	Family	Bundles	Fidelity	Pcap modality	Coverage
reth_pooledtx_decode_memory_amp	memory_amp	17	lab	none	attack; reth eth Transactions/PooledTransactions decoded into an unbounded Vec pre-#23718 → ~40× decode-memory amplification (436 MB per clean ≤10 MB message); saturating + distributed postures (RETH_TX_MEMORY_BUDGET; NR-2026-019)

reth_pooledtx_decode_memory_amp is a NullRabbit original-research finding (source_class: original, ground_truth_label: attack): pre-#23718 reth decodes the eth Transactions/PooledTransactions RLP list into an unbounded Vec before per-transaction validation, so a well-formed message that decodes cleanly (no misbehavior bump) allocates ~40× its wire size during decode — measured 436 MB for a ≤10 MB message, exceeding the 20 MB cap the fix introduces. Fixed in reth #23718 (paradigmxyz/reth). Captured against a protocol-compatible node (the eth/68 wire format is identical across clients).

Aptos (4 bundles)

Primitive	Family	Bundles	Fidelity	Pcap modality	Coverage
aptos_f10_modules_amp	response_amp	4	lab	none	attack; /v1/accounts/{addr}/modules?limit=200 REST egress amplification (~11,186×) (APT_F10; NR-2026-012); 4 postures

Cardano (6 bundles)

Primitive	Family	Bundles	Fidelity	Pcap modality	Coverage
cardano_submit_api_body_memory_pin	memory_amp	3	proxy	none	attack; cardano-submit-api has no Warp request-body cap → ~500 MB RSS pinned per request (CARDANO_SUBMIT_API_BODY; NR-2026-013); 3 postures
cardano_peershare_sort_cpu	compute_amp	3	proxy	none	attack; PeerSharing MsgShareRequest triggers an O(N log N) sort over 3000 peers with no rate limit → CPU pin (170µs/req) (H_CARDANO_4_PEERSHARING; NR-2026-017); 3 postures

XRP (3 bundles)

Primitive	Family	Bundles	Fidelity	Pcap modality	Coverage
rippled_batch_response_amp	response_amp	3	lab	none	attack; rippled JSON-RPC batch has no per-element cap → ~24.9× response amplification (F10_RIP_BATCH_AMP; NR-2026-014); 3 postures

Bitcoin (3 bundles)

Primitive	Family	Bundles	Fidelity	Pcap modality	Coverage
btc_bip324_prehandshake_ecdh_cpu	compute_amp	3	proxy	none	attack; Bitcoin Core BIP-324 v2 transport runs ellswift-ECDH+HKDF on the 64-byte inbound key before any auth/rate-limit → pre-auth CPU pin (55×/256× honest-peer latency) (BTC_V0_BIP324_PREHANDSHAKE_CPU; NR-2026-016); 3 postures

The 2026-07-06 additions (iota_f10_multiget_amp, sui_h02_state_sync_flood, aptos_f10_modules_amp, cardano_submit_api_body_memory_pin, rippled_batch_response_amp, btc_bip324_prehandshake_ecdh_cpu, cardano_peershare_sort_cpu) are NullRabbit original-research findings (source_class: original, ground_truth_label: attack), each the subject of a published operator advisory (NR-2026-010, NR-2026-011, NR-2026-012, NR-2026-013, NR-2026-014, NR-2026-016, NR-2026-017 respectively). They add Aptos, Cardano, XRP (XRPL), and Bitcoin as the fifth through eighth chains, and sui_h02_state_sync_flood populates gossip_abuse — the last previously-empty family — so all thirteen taxonomy families now carry bundles.

Posture variations capture distinct attack shapes against the same primitive (saturating throughput, low-volume, distributed-source, mimicry of organic traffic, historical-CVE patterns). Bundles within a primitive are intentionally non-identical so the dataset surfaces the variance the format is designed to represent.

Public-CVE replications (567 bundles)

These 567 bundles are faithful wire-level reproductions of 148 publicly disclosed CVEs / security advisories (across 113 attack primitives) in blockchain node software, captured against the vulnerable node so the attack signature is real. Each bundle's provenance.source_class is public-cve-replication and provenance.public_source links the original disclosure.

Bitcoin (106 bundles, 26 primitives)

Primitive	Family	Bundles	Public source
bitcoin_dup_input_crash	consensus_abuse	4	https://bitcoincore.org/en/2018/09/20/notice/
bitcoin_tx_relay_jamming	gossip_abuse	4	https://nvd.nist.gov/vuln/detail/CVE-2024-55563
btc_addr_overflow_flood	gossip_abuse	4	https://bitcoincore.org/en/2025/04/28/disclose-cve-2024-52919/
btc_alert_flood	memory_amp	4	https://nvd.nist.gov/vuln/detail/CVE-2016-10724
btc_blocktxn_double_fillblock	consensus_abuse	6	https://bitcoincore.org/en/2024/10/08/disclose-blocktxn-crash/
btc_bloom_divzero	compute_amp	4	https://nvd.nist.gov/vuln/detail/CVE-2013-5700
btc_cmpctblock_overflow	memory_amp	4	https://bitcoincore.org/en/2025/10/24/disclose-cve-2025-46597/
btc_cmpctblock_stall	connection_exhaustion	4	https://bitcoincore.org/en/2024/11/05/cb-stall-hindering-propagation/
btc_getdata_flood	connection_exhaustion	4	https://bitcoincore.org/en/2024/07/03/disclose-getdata-cpu/
btc_headers_genesis_spam	memory_amp	4	https://bitcoincore.org/en/2024/07/03/disclose-header-spam/
btc_headers_oom	memory_amp	4	https://bitcoincore.org/en/2024/09/18/disclose-headers-oom/
btc_inv_buffer_blowup	memory_amp	4	https://bitcoincore.org/en/2024/07/03/disclose-inv-buffer-blowup/
btc_inv_eviction_jam	gossip_abuse	4	https://bitcoincore.org/en/2024/07/03/disclose_already_asked_for/
btc_invalid_block_logfill	memory_amp	4	https://bitcoincore.org/en/2025/10/24/disclose-cve-2025-54605/
btc_invdos_flood	memory_amp	4	https://invdos.net/
btc_mutated_block	consensus_abuse	4	https://bitcoincore.org/en/2024/10/08/disclose-mutated-blocks-hindering-propagation/
btc_orphan_cpu	compute_amp	4	https://bitcoincore.org/en/2024/07/03/disclose-orphan-dos/
btc_oversized_recv_buffer	memory_amp	4	https://bitcoincore.org/en/2024/07/03/disclose_receive_buffer_oom/
btc_tx_maprelay	memory_amp	4	https://nvd.nist.gov/vuln/detail/CVE-2013-4627
btc_tx_quad_sighash	compute_amp	4	https://bitcoincore.org/en/2025/10/24/disclose-cve-2025-46598/
btc_version_selfnonce	memory_amp	4	https://bitcoincore.org/en/2025/10/24/disclose-cve-2025-54604/
btc_version_timestamp_overflow	consensus_abuse	4	https://bitcoincore.org/en/2024/07/03/disclose-timestamp-overflow/
p2p_getheaders_drain	response_amp	4	https://nvd.nist.gov/vuln/detail/CVE-2023-33297
sybil_peer_identity_flood	gossip_abuse	4	https://github.com/slowmist/Cryptocurrency-Security-Audit-Guide/blob/main/Blockchain-Common-Vulnerability-List.md
timejacking_peer_time_skew	consensus_abuse	4	https://github.com/slowmist/Cryptocurrency-Security-Audit-Guide/blob/main/Blockchain-Common-Vulnerability-List.md
tx_malleability_txid_mutate	consensus_abuse	4	https://github.com/slowmist/Cryptocurrency-Security-Audit-Guide/blob/main/Blockchain-Common-Vulnerability-List.md

Cosmos (98 bundles, 10 primitives)

Primitive	Family	Bundles	Public source
cometbft_bitarray_mismatch	consensus_abuse	16	https://github.com/cometbft/cometbft/security/advisories/GHSA-hrhf-2vcr-ghch
cometbft_blockpart_mismatch	consensus_abuse	16	https://github.com/advisories/GHSA-r3r4-g7hq-pq4f
cometbft_blocksync_height_decrease_stuck	memory_amp	4	CVE-2025-24371
cometbft_blocksync_malicious_peer	consensus_abuse	4	https://github.com/cometbft/cometbft/security/advisories/GHSA-hg58-rf2h-6rr7
cometbft_mempool_dup_tx_leak	memory_amp	4	https://github.com/cometbft/cometbft/security/advisories/GHSA-w24w-wp77-qffm
cometbft_voteext_panic	consensus_abuse	16	https://github.com/cometbft/cometbft/security/advisories/GHSA-p7mv-53f2-4cwj
cosmos_gogoproto_skippy	compute_amp	16	https://osv.dev/vulnerability/CVE-2021-3121
cosmos_group_divzero_halt	consensus_abuse	8	https://github.com/cosmos/cosmos-sdk/security/advisories/GHSA-x5vx-95h7-rv4p
cosmos_p2p_conn_flood	connection_exhaustion	12	https://github.com/tendermint/tendermint/security/advisories/GHSA-v24h-pjjv-mcp6
cosmos_protobuf_nest_bomb	compute_amp	2	https://github.com/cosmos/cosmos-sdk/security/advisories/GHSA-8wcc-m6j2-qxvm

Ethereum (96 bundles, 18 primitives)

Primitive	Family	Bundles	Public source
discv4_findnode_amplification	memory_amp	4	NethermindEth/nethermind#12211 — 'Resolve some discovery dos vectors (which hive revealed)': discv4 (UDP node-discovery) DoS mitigation — endpoint-proof bonding tied to the exact UDP IP:port before honouring FINDNODE, plus a 50/50 request/response rate-limit split so NEIGHBORS response bursts can't starve legitimate queries. https://github.com/NethermindEth/nethermind/pull/12211
geth_alien_peer_pool_pollution	connection_exhaustion	4	https://github.com/slowmist/Cryptocurrency-Security-Audit-Guide/blob/main/Blockchain-Common-Vulnerability-List.md
geth_auth_zero_pubkey	compute_amp	4	https://github.com/ethereum/go-ethereum/security/advisories/GHSA-q26p-9cq4-7fc2
geth_blob_kzg_dos	compute_amp	4	https://github.com/ethereum/go-ethereum/security/advisories/GHSA-mq3p-rrmp-79jg
geth_devp2p_ping_flood	connection_exhaustion	4	https://github.com/ethereum/go-ethereum/security/advisories/GHSA-ppjg-v974-84cm
geth_eth_receipt_flood	memory_amp	12	https://reports.immunefi.com/ethereum-protocol-or-attackathon/37466-bc-medium-evil-client-oom-crash-fast-p2p-crash
geth_ethash_memory_exhaustion_dos	memory_amp	4	CVE-2021-42219 — https://nvd.nist.gov/vuln/detail/CVE-2021-42219 (MITRE: https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2021-42219). go-ethereum v1.10.9 remote DoS: an excessive amount of block/header messages to a node drives repeated ethash cache generation in consensus/ethash/algorithm.go -> memory exhaustion. CVSS 3.1 7.5 (AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H). No vendor GHSA (third-party exploit-doc ref).
geth_getblockheaders_count_zero	memory_amp	12	https://github.com/ethereum/go-ethereum/security/advisories/GHSA-4xc9-8hmq-j652
geth_graphql_aliased_logs_dos	memory_amp	4	https://github.com/advisories/GHSA-v9jh-j8px-98vq
geth_les_getproofsv2_dos	compute_amp	4	https://nvd.nist.gov/vuln/detail/CVE-2020-26264
geth_les_skip_negative	compute_amp	4	https://peckshield.medium.com/epod-ethereum-packet-of-death-cve-2018-12018-fc9ee944843e
geth_mempool_spend_all_flood	memory_amp	4	https://nvd.nist.gov/vuln/detail/CVE-2022-23328
geth_rlpx_auth_flood	connection_exhaustion	4	https://notes.ethereum.org/gDWKW5RtSym02t2aGYkmSQ
geth_rpc_unlocked_wallet_drain	auth_bypass	4	https://github.com/slowmist/Cryptocurrency-Security-Audit-Guide/blob/main/Blockchain-Common-Vulnerability-List.md
geth_snap_trienode_dos	compute_amp	12	https://github.com/ethereum/go-ethereum/security/advisories/GHSA-59hh-656j-3p7v
geth_sybil_identity_flood	gossip_abuse	4	https://github.com/slowmist/Cryptocurrency-Security-Audit-Guide/blob/main/Blockchain-Common-Vulnerability-List.md
geth_tcp_handshake_flood	connection_exhaustion	4	https://reports.immunefi.com/ethereum-protocol-or-attackathon/37120-bc-insight-remote-handshake-based-tcp-30303-flooding-leads-to-an-out-of-memory-crash
geth_tx_future_flood	memory_amp	4	https://github.com/advisories/GHSA-pvx3-gm3c-gmpr

quic (68 bundles, 17 primitives)

Primitive	Family	Bundles	Public source
coredns_doq_stream_slowloris	connection_exhaustion	4	CVE-2025-47950 / GHSA-cvx7-x8pj-x2gw — CoreDNS DNS-over-QUIC (DoQ) memory exhaustion via stream amplification (CoreDNS < 1.12.2): the DoQ server spawns a goroutine per QUIC stream with no stream/goroutine limit and reads the DoQ length prefix + body with blocking io.ReadFull() calls that have no per-stream read deadline; opening many streams and sending 1 byte on each pins goroutines -> OOM. https://github.com/coredns/coredns/security/advisories/GHSA-cvx7-x8pj-x2gw
dnsdist_doq_concurrent_conn_exhaustion	connection_exhaustion	4	CVE-2026-33254 / PowerDNS Security Advisory 2026-04 (dnsdist) — PowerDNS dnsdist DoQ/DoH3 concurrent-connection memory exhaustion: the DNS-over-QUIC (addDOQLocal) and DNS-over-HTTP/3 (addDOH3Local) frontends commit per-connection state with no cap on the number of concurrent DoQ/DoH3 connections, so an attacker opening many concurrent connections drives unbounded memory growth -> DoS (CWE-770). DoQ/DoH3 are off by default (CVSS 5.3). https://dnsdist.org/security-advisories/powerdns-advisory-2026-04.html
dnsdist_doq_error_query_mem_leak	memory_amp	4	CVE-2026-33595 / PowerDNS Security Advisory 2026-04 — PowerDNS dnsdist DoQ/DoH3 memory exhaustion via error-triggering queries: over a single DNS-over-QUIC or DNS-over-HTTP3 connection an attacker sends a large run of error-triggering DNS queries; dnsdist attaches per-error bookkeeping to the connection and does not free it until the connection closes, so memory accumulates without bound -> DoS. CWE-400; CVSS 3.1 Medium (~5.3) (CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:L). DoQ/DoH3 listeners are OFF BY DEFAULT.
lsquic_initial_prehandshake_leak	memory_amp	4	CVE-2025-54939 (QUIC-LEAK) — Imperva disclosure (https://www.imperva.com/blog/quic-leak-cve-2025-54939-new-high-risk-pre-handshake-remote-denial-of-service-in-lsquic-quic-implementation/); LiteSpeed advisory https://blog.litespeedtech.com/2025/08/18/litespeed-security-update/. LSQUIC ('LiteSpeed QUIC (LSQUIC) Library LSQUIC Engine Packet In Memory Leak') coalesces multiple QUIC packets per UDP datagram and allocates one 96-byte packet_in per coalesced packet; when trailing coalesced Initial packets fail DCID validation, lsquic_engine.c frees only the first packet's packet_in and never calls lsquic_mm_put_packet_in for the rest -> pre-handshake memory leak that bypasses all post-handshake connection/stream/flow limits -> OOM (70% of bandwidth).
msquic_dup_tp_versioninfo_leak	memory_amp	4	CVE-2024-26190 / GHSA-2x7m-gf85-3745 — microsoft/msquic: 'Remote Denial of Service Vulnerability' — memory leak from multiple decodes of the QUIC transport parameters TLS extension. A remote unauthenticated peer sends a ClientHello whose TLS extension list contains multiple duplicate quic_transport_parameters extensions (ext type 0x0039), each carrying a version_information transport parameter (TP id 0x11); QuicCryptoTlsReadExtensions did not reject the duplicate, so QuicCryptoTlsDecodeTransportParameters ran once per occurrence, each CXPLAT_ALLOC'ing a fresh VersionInfo (QUIC_POOL_VERSION_INFO) and then zeroing the struct without freeing the prior pointer -> one leaked allocation per duplicate. 'The MsQuic server will continue to leak memory until no more is available, resulting in a denial of service.' CVSS 3.1 7.5 (High; CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H). Affected msquic < 2.1.12 (2.1.x), < 2.2.7 (2.2.x), < 2.3.5; fixed 2.1.12, 2.2.7, 2.3.5 (commit d364feeda0dd). Advisory: https://github.com/microsoft/msquic/security/advisories/GHSA-2x7m-gf85-3745
quic_conn_id_memory_exhaustion	memory_amp	4	CVE-2024-22189 / GHSA-c33x-xqrf-c478 — quic-go < 0.42.0: 'QUIC's Connection ID Mechanism vulnerable to Memory Exhaustion Attack'. https://github.com/quic-go/quic-go/security/advisories/GHSA-c33x-xqrf-c478
quic_go_path_challenge_flood	memory_amp	4	CVE-2023-49295 / GHSA-ppxx-5m9h-6vxf — quic-go/quic-go: 'Memory Exhaustion Attack against QUIC's Path Validation Mechanism.' An attacker floods the peer with a large number of PATH_CHALLENGE frames (RFC 9000 §8.2 requires a PATH_RESPONSE for each), then prevents those responses from being sent by collapsing the peer's congestion window (selective ACK) and manipulating its RTT estimate — so the un-sendable PATH_RESPONSE frames queue internally without bound, exhausting memory (availability DoS, severity Moderate). No workaround; the fix bounds retained path-validation state. Affected quic-go v0.40.0, <= v0.39.3, <= v0.38.1, <= v0.37.6; fixed v0.40.1, v0.39.4, v0.38.2, v0.37.7. Reported by marten-seemann. Advisory: https://github.com/quic-go/quic-go/security/advisories/GHSA-ppxx-5m9h-6vxf
quic_go_qpack_decompression_bomb	memory_amp	4	CVE-2025-64702 / GHSA-g754-hx8w-x2g6 — quic-go/quic-go: 'HTTP/3 QPACK Header Expansion Memory Exhaustion.' quic-go's HTTP/3 decoder limits the ENCODED HEADERS frame size (1 MB (http3.Server.MaxHeaderBytes) server / 10 MB (http3.Transport.MaxResponseHeaderBytes) client) but not the DECODED field-section size. A crafted QPACK HEADERS block that densely references long static-table entries decompresses to ~50x its encoded size; RFC 9114 requires enforcing SETTINGS_MAX_FIELD_SECTION_SIZE, which quic-go <= v0.56.0 did not, so the decoder allocates the fully-expanded field section unbounded -> memory exhaustion (availability DoS, severity Moderate, CVSS 5.3 CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:L). Fixed v0.57.0 (SETTINGS_MAX_FIELD_SECTION_SIZE + incremental QPACK decoding, abort early). Affected quic-go <= v0.56.0. Reported by sfoxio. Advisory: https://github.com/quic-go/quic-go/security/advisories/GHSA-g754-hx8w-x2g6
quic_optimistic_ack_cwnd_growth	memory_amp	4	CVE-2025-4820 / GHSA-2v9p-3p3h-w56j — https://github.com/cloudflare/quiche/security/advisories/GHSA-2v9p-3p3h-w56j (see also https://blog.cloudflare.com/defending-quic-from-acknowledgement-based-ddos-attacks/ and RFC 9000 §21.4 Optimistic ACK Attack). cloudflare/quiche < 0.24.4 incorrectly grows its congestion window in response to ACK frames that acknowledge packets the peer never received (optimistic / opportunistic ACK), allowing more bytes in flight than the path supports and transmission faster than the path can handle. Fixed in quiche 0.24.4 (dynamic CWND-aware packet-number skipping so a forward ACK for a not-yet-sent packet number is caught).
quic_scid_collision_hash_dos	compute_amp	4	CVE-2025-47200 — NCC Group / Fox-IT technical advisory 'Hash Denial-of-Service Attack in Multiple QUIC Implementations'. QUIC servers index the connection table by the peer-chosen Source Connection ID (SCID, 8..20 bytes). Implementations hashing the SCID with a weak non-seeded function (FNV, xxHash, or a multiplicative hash = hash*31 + char) let an attacker pre-compute DISTINCT SCIDs that all collide into one bucket; opening many such connections degrades table insert/lookup from O(1) to O(n) (amortised O(n^2)), burning server CPU. Only SipHash resists it. Affected: xquic <= 1.8.1 (fixed 1.8.2), Apache Traffic Server <= 10.0.3 / 9.2.x <= 9.2.8, ngtcp2 1.10.0 example server (all CVE-2025-47200); kwik <= 0.10.0 (CVE-2025-23020, fixed 0.10.1); lsquic <= 4.0.12 (CVE-2025-24947, fixed 4.2.0); picoquic commit 3827b00 (CVE-2025-24946, fixed b80fd3f). https://www.nccgroup.com/research/technical-advisory-hash-denial-of-service-attack-in-multiple-quic-implementations/
quiche_ack_never_sent_cwnd	memory_amp	4	CVE-2025-4821 / GHSA-6m38-4r9r-5c4m — https://github.com/cloudflare/quiche/security/advisories/GHSA-6m38-4r9r-5c4m (NVD: https://nvd.nist.gov/vuln/detail/CVE-2025-4821 ; see also https://blog.cloudflare.com/defending-quic-from-acknowledgement-based-ddos-attacks/ and RFC 9000 §19.3/§21.4). cloudflare/quiche < 0.24.4 incorrectly grows its congestion window in response to ACK frames covering a large range of packet numbers INCLUDING packet numbers that were NEVER SENT (invalid ACK ranges) — allowing more bytes in flight than the path supports and, in the extreme case, growing the window beyond the limit of the internal variable's type, causing an OVERFLOW PANIC. Fixed in quiche 0.24.4 (proper ACK-range validation rejecting ranges over never-sent packet numbers).
quiche_conn_id_retirement_unbounded	memory_amp	4	CVE-2024-1410 / GHSA-xhg9-xwch-vr7x — Cloudflare quiche (< 0.19.2; >= 0.20.0, < 0.20.1): 'quiche vulnerable to unbounded storage of information related to connection ID retirement'. https://github.com/advisories/GHSA-xhg9-xwch-vr7x
quiche_crypto_frame_flood	memory_amp	4	CVE-2024-1765 / GHSA-78wx-jg4j-5j6g — cloudflare/quiche: 'Unlimited resource allocation by QUIC CRYPTO frames flooding.' A remote peer repeatedly sends an unlimited number of QUIC CRYPTO frames (type 0x06) over an established connection; quiche buffered the CRYPTO reassembly stream without bounding total retained state, so the flood makes memory usage of the quiche server OR client climb for the duration of the connection (availability DoS, CVSS 3.1 5.9). Affected quiche <= 0.19.1, 0.20.0; fixed in 0.19.2, 0.20.1. Reported by Marten Seemann. Advisory: https://github.com/cloudflare/quiche/security/advisories/GHSA-78wx-jg4j-5j6g
quiche_path_challenge_queue	memory_amp	4	CVE-2023-6193 / GHSA-w3vp-jw9m-f9pm — cloudflare/quiche: 'Unbounded queuing of path validation messages.' QUIC path validation (RFC 9000 §8.2) makes the recipient of a PATH_CHALLENGE frame (type 0x1a) reply with a PATH_RESPONSE (type 0x1b) echoing the 8-byte challenge. An unauthenticated remote attacker sends PATH_CHALLENGE frames — changing source address so each path migration forces path validation — and manipulates the connection (e.g. restricting the peer's congestion window) so PATH_RESPONSEs can only be sent slower than PATH_CHALLENGEs are received; quiche stores the path-validation data in an UNBOUNDED queue, so memory grows without limit for the connection's lifetime (availability DoS, CWE-400, CVSS 3.1 5.3 Moderate). Affected quiche 0.15.0 through 0.19.0; fixed in 0.19.1. Advisory: https://github.com/cloudflare/quiche/security/advisories/GHSA-w3vp-jw9m-f9pm
quinn_gap_fragment_reassembly_oom	memory_amp	4	GHSA-4w2j-m93h-cj5j (https://github.com/advisories/GHSA-4w2j-m93h-cj5j) / RUSTSEC-2026-0185 (https://rustsec.org/advisories/RUSTSEC-2026-0185.html) — quinn-proto: remote memory exhaustion from unbounded out-of-order stream reassembly. The per-stream Assembler retains out-of-order received STREAM data until a contiguous prefix can be delivered; a peer sending many small STREAM fragments at gapped (non-contiguous) offsets forces un-coalescible entries whose per-fragment overhead grows without bound while never becoming deliverable.
s2n_quic_crypto_offset_amplification	memory_amp	4	CVE-2026-10740 / GHSA-9q54-f358-3fqf — aws/s2n-quic: CRYPTO-frame offset amplification. The CRYPTO-frame reassembly buffer lacks maximum-size enforcement, so a CRYPTO frame (type 0x06) carrying an artificially HIGH offset forces the receiver to reserve reassembly buffer up to that offset even though the delivered body is tiny: a single 1200-byte packet causes ~9.4 MB of allocation (7800x). No authentication or valid handshake is required — the crafted CRYPTO frames ride pre-handshake QUIC Initial packets, so an unauthenticated peer can repeatedly transmit them to exhaust server memory (availability DoS, CVSS 3.1 5.3). Affected s2n-quic <= 1.81.0; fixed in 1.82.0. Advisory: https://github.com/aws/s2n-quic/security/advisories/GHSA-9q54-f358-3fqf
s2n_quic_stream_limit_exhaustion	connection_exhaustion	4	GHSA-475v-pq2g-fp9g — AWS s2n-quic <= v1.30.0: 's2n-quic potential denial of service via crafted stream frames' — uncontrolled stream-limit increase when closing remote-initiated streams. https://github.com/aws/s2n-quic/security/advisories/GHSA-475v-pq2g-fp9g

Monero (50 bundles, 3 primitives)

Primitive	Family	Bundles	Public source
monero_levin_array_memcorrupt	compute_amp	2	https://www.talosintelligence.com/vulnerability_reports/TALOS-2018-0637
monero_portable_storage_oom	memory_amp	32	https://github.com/monero-project/monero/pull/7190
monero_rpc_conn_exhaustion	connection_exhaustion	16	https://nvd.nist.gov/vuln/detail/CVE-2025-26819

libp2p (44 bundles, 11 primitives)

Primitive	Family	Bundles	Public source
go_libp2p_oversized_rsa_key_cpu_burn	compute_amp	4	https://github.com/libp2p/go-libp2p/security/advisories/GHSA-876p-8259-xjgg
gossipsub_ihave_iwant_flood	gossip_abuse	4	https://nvd.nist.gov/vuln/detail/CVE-2026-49866
gossipsub_prune_backoff_overflow	gossip_abuse	4	https://github.com/libp2p/rust-libp2p/security/advisories/GHSA-xqmp-fxgv-xvq5
gossipsub_subscribe_flood	gossip_abuse	4	https://github.com/advisories/GHSA-4f8r-922h-2vgv
libp2p_conn_resource_exhaustion	connection_exhaustion	4	https://github.com/libp2p/js-libp2p/security/advisories/GHSA-f44q-634c-jvwv
libp2p_dht_putvalue_disk_exhaustion	memory_amp	4	https://github.com/advisories/GHSA-32mq-hpph-xfvr
libp2p_rendezvous_cookie_exhaustion	memory_amp	4	https://github.com/advisories/GHSA-v5hw-cv9c-rpg7
libp2p_rendezvous_namespace_oom	memory_amp	4	https://github.com/advisories/GHSA-cqfx-gf56-8x59
libp2p_signed_peer_record_flood	gossip_abuse	4	https://github.com/advisories/GHSA-gcq9-qqwx-rgj3
libp2p_stream_exhaustion	connection_exhaustion	4	https://github.com/advisories/GHSA-j7qp-mfxf-8xjw
yamux_ping_flood_pending_frames_oom	memory_amp	4	https://github.com/libp2p/rust-yamux/security/advisories/GHSA-3999-5ffv-wp2r

zcash (32 bundles, 8 primitives)

Primitive	Family	Bundles	Public source
zcash_sapling_woodchip_tx_flood	gossip_abuse	4	https://nvd.nist.gov/vuln/detail/CVE-2019-11636
zcash_zebra_addr_vector_preallocation_amplification	memory_amp	4	https://github.com/advisories/GHSA-xr93-pcq3-pxf8
zcash_zebra_coinbase_script_preallocation_amplification	memory_amp	4	https://nvd.nist.gov/vuln/detail/CVE-2026-44500
zcash_zebra_equihash_solution_preallocation_amplification	memory_amp	4	https://nvd.nist.gov/vuln/detail/CVE-2026-44500
zcash_zebra_headers_message_preallocation_amplification	memory_amp	4	https://nvd.nist.gov/vuln/detail/CVE-2026-44500
zebra_block_discovery_dos	gossip_abuse	4	CVE-2026-44499 / GHSA-h9hm-m2xj-4rq9 — https://github.com/ZcashFoundation/zebra/security/advisories/GHSA-h9hm-m2xj-4rq9 (also https://nvd.nist.gov/vuln/detail/CVE-2026-44499). Permanent Block Discovery Halt in Zebra via Gossip Queue Saturation and Syncer Poisoning.
zebra_rpc_premature_disconnect_crash	connection_exhaustion	4	CVE-2026-41585 / GHSA-29x4-r6jv-ff4w https://github.com/ZcashFoundation/zebra/security/advisories/GHSA-29x4-r6jv-ff4w (Zebra JSON-RPC HTTP middleware: DoS via interrupted/truncated request body from an authenticated client)
zebra_sync_restart_poisoning	gossip_abuse	4	CVE-2026-52737 / GHSA-gvjc-3w7c-92jx — https://github.com/ZcashFoundation/zebra/security/advisories/GHSA-gvjc-3w7c-92jx

http2 (24 bundles, 6 primitives)

Primitive	Family	Bundles	Public source
envoy_http2_hpack_cookie_amplification	memory_amp	4	CVE-2026-47774 / GHSA-22m2-hvr2-xqc8 (https://github.com/envoyproxy/envoy/security/advisories/GHSA-22m2-hvr2-xqc8). Envoy HTTP/2 codec (oghttp2/quiche): HPACK header-block limits enforced on ENCODED bytes with no total-DECODED-size limit, plus cookie bytes escaping max_request_headers_kb accounting. Affected Envoy < 1.39; patched 1.35.11 / 1.36.7 / 1.37.3 / 1.38.1. CWE-405 + CWE-770; CVSS 7.5.
http2_bomb_indexed_ref_window_pin	memory_amp	4	CVE-2026-49975 ("HTTP/2 Bomb"). HTTP/2 servers (nginx <1.29.8, Apache httpd mod_http2 <2.0.41, Envoy <=1.37.2, Cloudflare Pingora) enforce HPACK limits on ENCODED header-block bytes without bounding the total DECODED header list, and hold each stream's amplified decoded allocation until the stream makes flow-control progress. An attacker seeds the HPACK dynamic table with one generic header then emits thousands of 1-byte indexed references (0x80
http2_continuation_flood	compute_amp	4	https://github.com/advisories/GHSA-w9fj-cfpg-grvv
http2_madeyoureset_flood	connection_exhaustion	4	CVE-2025-8671 ("MadeYouReset") / CVE-2025-55163 (GHSA-prj3-ccx8-p6x4) / CERT/CC VU#767506 — https://kb.cert.org/vuls/id/767506 , https://github.com/netty/netty/security/advisories/GHSA-prj3-ccx8-p6x4 . HTTP/2 DoS in which an attacker sends protocol-valid-but-misused control frames (WINDOW_UPDATE increment 0, HEADERS/DATA on a half-closed stream, or PRIORITY with a bad length) that force the SERVER to emit RST_STREAM. The reset stream immediately stops counting against MAX_CONCURRENT_STREAMS (typ. 100) yet the backend keeps processing the request -> unbounded concurrent work -> resource exhaustion (OOM/CPU) -> DoS. Umbrella CVSS 3.1 7.5 HIGH; Netty CVE-2025-55163 CVSS 4.0 8.2 HIGH (patched 4.1.124.Final / 4.2.4.Final). DISTINCT from CVE-2023-44487 Rapid Reset (client-emitted RST): here the RST is server-emitted so client-side rapid-reset mitigations do not apply. Discovered by Tel Aviv University + Imperva. Affects Netty, Apache Tomcat (CVE-2025-48989), Eclipse Jetty (CVE-2025-5115), F5 BIG-IP (CVE-2025-54500), gRPC and many other HTTP/2 stacks (CERT/CC VU#767506).
http2_malformed_priority_leak_flood	memory_amp	4	CVE-2025-31650 (Apache Tomcat) — https://nvd.nist.gov/vuln/detail/CVE-2025-31650 , https://lists.apache.org/thread/j6zzk0y3yym9pzfzkq5vcyxzz0yzh826 . Improper Input Validation / Incomplete Cleanup (CWE-459): incorrect error handling for some INVALID HTTP/2 priority values resulted in incomplete clean-up of the failed request, creating a memory leak; a large number of such requests triggers an OutOfMemoryError → DoS. Vulnerable path: Http2Parser.readPriorityUpdateFrame (and Stream.emitHeader) → Priority.parsePriority(r) throws IllegalArgumentException on an invalid RFC 9218 priority value (e.g. out-of-range urgency u=99), previously un-caught. CVSS 3.1 7.5 HIGH (AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H). Affected: Apache Tomcat 9.0.76-9.0.102, 10.1.10-10.1.39, 11.0.0-M2-11.0.5 (and EOL-at-disclosure 8.5.90-8.5.100); fixed 9.0.104 / 10.1.40 / 11.0.6. Fixed by wrapping Priority.parsePriority in try/catch + Http2UpgradeHandler.upgradeDispatch cleanup.
http2_settings_ack_stream_cap_bypass	connection_exhaustion	4	CVE-2025-53506 (GHSA-25xr-qj8w-c4vf) — https://github.com/advisories/GHSA-25xr-qj8w-c4vf , https://www.openwall.com/lists/oss-security/2025/07/10/13 . Uncontrolled Resource Consumption in Apache Tomcat Coyote HTTP/2: at connection start the server's initial SETTINGS frame reduces SETTINGS_MAX_CONCURRENT_STREAMS to the configured cap (default 100), but Tomcat did not ENFORCE that reduced cap until the client acknowledged the SETTINGS frame (a SETTINGS frame with the ACK flag). A remote unauthenticated client that never sends the ACK immediately opens far more concurrent streams than the cap; Coyote accepts every stream, allocates a request/response processor + buffers per stream, never emits REFUSED_STREAM / GOAWAY, and exhausts memory and threads -> DoS. CVSS 3.1 7.5 HIGH (CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H), CWE-400. Affected: Apache Tomcat 11.0.0-M1..11.0.8; 10.1.0-M1..10.1.42; 9.0.0.M1..9.0.106; 8.5.0..8.5.100 (EOL) (patched 11.0.9 / 10.1.43 / 9.0.107). DISTINCT from CVE-2023-44487 Rapid Reset (client RST_STREAM flood), CVE-2025-8671 MadeYouReset (server-emitted RST via a misuse frame) and CVE-2026-33871 CONTINUATION flood (single-stream header-decode pin): this attack sends NO RST_STREAM and NO misuse/CONTINUATION frame — it opens many complete concurrent streams past an un-enforced cap.

Solana (17 bundles, 5 primitives)

Primitive	Family	Bundles	Public source
firedancer_metrics_flood_crash	connection_exhaustion	4	https://github.com/firedancer-io/firedancer/releases/tag/v0.106.11814 — Firedancer v0.106.11814 (Testnet) Bug Fixes: "Fixed an issue where a connection flood to the metrics server could crash the validator." Firedancer's fd_metrics tile serves Prometheus-compatible metrics over HTTP (default 127.0.0.1:7999/metrics; listen address operator-configurable / internet-exposable per the Firedancer bug-bounty guidance). A remote inbound TCP connection flood to the metrics port could crash the whole validator prior to this fix.
sol_tpu_quic_handshake_flood	connection_exhaustion	3	https://neodyme.io/reports/Firedancer-v0.4.pdf
sol_tpu_quic_initial_cpu	connection_exhaustion	3	https://neodyme.io/reports/Firedancer.pdf
sol_tpu_quic_slowloris	connection_exhaustion	3	https://neodyme.io/reports/Firedancer-v0.4.pdf
solana_repair_protocol_dos	compute_amp	4	https://github.com/solana-labs/solana/releases/tag/v1.1.16 — release note "Avoid possible repair orphan DoS"; fix PR https://github.com/solana-labs/solana/pull/10290 ("Fix run_orphan DOS", merged 2020-05-28), core/src/serve_repair.rs.

Sui (12 bundles, 3 primitives)

Primitive	Family	Bundles	Public source
sui_disassemble_panic	compute_amp	4	https://medium.com/certik-skyfall/blockchain-rpc-vulnerabilities-why-memory-safe-blockchain-rpc-nodes-are-not-panic-free-9fbb990115e0
sui_move_recursion	compute_amp	4	https://nvd.nist.gov/vuln/detail/CVE-2023-36184
sui_verifier_hamsterwheel	compute_amp	4	https://medium.com/certik-skyfall/the-hamsterwheel-an-in-depth-exploration-of-a-novel-attack-vector-on-the-sui-blockchain-522f80623bc7

http3 (8 bundles, 2 primitives)

Primitive	Family	Bundles	Public source
envoy_http3_qpack_blocked_decode_leak	memory_amp	4	GHSA-p7c7-7c47-pwch — envoyproxy/envoy: 'HTTP/3 QPACK blocked-decoding memory exhaustion.' Envoy's HTTP/3 QPACK decoder mis-accounts QUIC flow control for BLOCKED header blocks: when a received field section references dynamic-table entries with a Required Insert Count higher than the entries actually inserted (the encoder-stream inserts have not arrived), the block cannot be decoded and is buffered on the decoder heap, yet 'stream-level and connection-level flow-control windows are returned to the peer even though the payload bytes are still resident in heap memory.' The attacker never sends the referenced inserts (blocks never unblock) and, because the credit was returned, progressively feeds more HEADERS payload bytes into this uncontrolled internal buffer, circumventing QUIC flow-control limits -> unbounded memory growth and DoS of the HTTP/3 stack. Severity High, CVSS 7.5 (CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H); no CVE assigned. Affected Envoy < 1.39; fixed 1.35.13, 1.36.9, 1.37.5, 1.38.3. No auth/privileges required. Advisory: https://github.com/envoyproxy/envoy/security/advisories/GHSA-p7c7-7c47-pwch
netty_http3_field_section_oom	memory_amp	4	CVE-2026-44892 / GHSA-c2rx-5r8w-8xr2 — netty/netty (io.netty:netty-codec-http3 (Maven)): 'Vulnerable Default Configuration Which Leads to Denial of Service via Unbounded HTTP/3 Header Size.' Netty enforces safe 8192-byte header limits for HTTP/1.1 and HTTP/2, but the HTTP/3 Http3ConnectionHandler follows RFC 9114's unlimited default: when a peer omits HTTP3_SETTINGS_MAX_FIELD_SECTION_SIZE, maxHeaderListSize defaults to the virtually-unbounded (1L << 62) - 1. A peer that omits the setting and sends a HEADERS block with an enormous number of header fields makes the codec buffer field data until the JVM exhausts heap -> OutOfMemoryError (availability DoS, severity High, CVSS 7.5 CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H; CWE-400 (Uncontrolled Resource Consumption); CWE-1188 (Insecure Default Initialization)). Fixed 4.2.15.Final (default replaced with a hardcoded 8192-byte cap Http3CodecUtils.DEFAULT_MAX_FIELD_SECTION_SIZE). Affected >= 4.2.0.Final, <= 4.2.14.Final. Advisory: https://github.com/netty/netty/security/advisories/GHSA-c2rx-5r8w-8xr2

polkadot (4 bundles, 1 primitive)

Primitive	Family	Bundles	Public source
substrate_h2_rapid_reset	memory_amp	4	CVE-2023-26964 (RUSTSEC-2023-0034 / GHSA-f8vr-r385-rh5r) — https://nvd.nist.gov/vuln/detail/CVE-2023-26964, https://rustsec.org/advisories/RUSTSEC-2023-0034.html . HTTP/2 Rapid Reset (CVE-2023-44487 class) resource-exhaustion DoS in the Rust h2 crate < 0.3.17: a flood of HEADERS/RST_STREAM frames grows the pending-accept-reset stream queue unbounded in memory -> OOM. CVSS 3.1 7.5 HIGH (AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H), CWE-770. Consumed by substrate/polkadot via hyper on the RPC & prometheus HTTP servers; bumped 0.3.16 -> 0.3.17 in paritytech/substrate#13915 (https://github.com/paritytech/substrate/pull/13915). Fix: hyperium/h2#668 (https://github.com/hyperium/h2/pull/668) adds max_pending_accept_reset_streams + GOAWAY(ENHANCE_YOUR_CALM).

ipfs (4 bundles, 1 primitive)

Primitive	Family	Bundles	Public source
ipfs_dht_sybil_censorship	gossip_abuse	4	https://github.com/advisories/GHSA-mqr9-hjr8-2m9w

Litecoin (2 bundles, 1 primitive)

Primitive	Family	Bundles	Public source
p2p_getheaders_drain	response_amp	2	https://nvd.nist.gov/vuln/detail/CVE-2023-33297

Dogecoin (2 bundles, 1 primitive)

Primitive	Family	Bundles	Public source
p2p_getheaders_drain	response_amp	2	https://nvd.nist.gov/vuln/detail/CVE-2023-33297

How the corpus is built

NullRabbit operates a reproducer pipeline that captures bundles under controlled conditions: validator daemons run on isolated infrastructure, workloads (attack or benign) are scripted and replayable, capture is synchronous across all five modalities, and every bundle's provenance is hashed end-to-end.

The full archived corpus (corpus_v1.0 through corpus_v1.10) contains thousands of bundles across additional primitives, families, and fidelity classes. This release is a curated three-hundred-and-ninety-five-bundle subset selected to demonstrate format coverage and provide training material for external researchers, with the rest of the corpus retained as the proprietary training surface for NullRabbit's production models.

The corpus is versioned and immutable. Each version archives only bundles new to that version. Failed iterations stay on disk as evidence; only versions passing close-out are promoted to archive.

Methodology

NullRabbit's ML work against this corpus follows pre-registration discipline borrowed from the empirical-reproducibility literature. Every training cycle has a design document committed before the trainer runs. Audits run on close against sanity floors, per-feature audit trails, and falsification holdouts. Where an audit fires, training halts, the design is re-registered, and the prior version is retracted in writing.

The methodology, the iterative leak-surface peeling pattern, and the multi-cycle audit arc are documented in the substrate paper (in preparation) and in companion docs in the NullRabbit Labs research track. The corpus format and family taxonomy are open; the methodology is open; the specific corpus contents beyond this public sample are proprietary.

The governance framework for autonomous defensive systems built on this data is published separately as the earned-autonomy paper (Zenodo DOI 10.5281/zenodo.18406828).

Intended uses

This dataset is intended for:

• Research on multi-modal anomaly detection against validator infrastructure, particularly cross-chain generalisation claims.
• Benchmarking for adversarial-robustness work on network-layer ML detection.
• Format adoption: as the reference public sample of the bundle v1 format. Anyone building their own corpus against nr-bundle-spec can use these bundles as ground-truth examples.
• Methodology study: the bundles are the data layer; the pre-registration / falsifiability methodology applied to training against them is documented in the substrate paper.

Limitations

• Lab fidelity only. All bundles in this release were captured in a controlled reproducer environment, not on mainnet or testnet. The capture pipeline has known fidelity envelopes; production-deployment generalisation is an open empirical question.
• Thirteen chains and stacks. Sui, Solana, IOTA, Cosmos, Aptos, Cardano, XRP (XRPL), Bitcoin, Ethereum, Monero, Dogecoin, Litecoin, and libp2p. Cross-chain claims in published work apply specifically to this corpus; the measured held-out-chain transfer figures above predate the public-CVE-replication additions and have not been re-measured on the full 13-chain corpus.
• Curated subset. Three hundred and ninety-five bundles do not represent the full distribution of attack shapes in NullRabbit's archived corpus. The subset is selected for coverage, not for statistical sufficiency. Training a production-grade model on this subset alone is not recommended.
• Raw packets.pcap is dropped from every bundle in this release. BundleFiles.packets_pcap=False is set in every manifest and no pre-termination pcap variant is retained (dropped via tools/strip_pcap.py from nr-bundle-spec). The remaining Parquet modalities (host plus whichever of app, protocol, responses were captured for the workload) stay intact for all three hundred and ninety-five bundles. Future dataset versions may include a header-only stripped variant of packets.pcap for lab bundles once payload-stripping tooling has been verified.
• Modality presence varies by workload. Passive-workload benigns (sui_BENIGN_passive_fullnode, solana_BENIGN_validator_passive) do not serve RPC traffic during capture; their responses.parquet is absent or zero-rows and BundleFiles.responses_parquet=False in the manifest. Consumers must check the manifest's files block before reading each modality. vectors.parquet is reserved-but-unpopulated in this release; consumers derive feature vectors at inference time.
• Disclosure status varies by primitive. SOL_F10/F14/P07 are publicly disclosed per NR-2026-001. The other original-research primitives each carry a published operator advisory (NR-2026-003 through NR-2026-017). The public-cve-replication primitives reproduce already-public CVEs / security advisories; each links its original disclosure via provenance.public_source (also tabulated above).

How to use this dataset

Each bundle is a directory of Parquet files plus a manifest, with schemas that vary by modality. The dataset is tree-shaped, not tabular, so the canonical way to consume it is snapshot_download followed by the bundle-spec reference parsers:

from huggingface_hub import snapshot_download
from bundle_spec import BundleManifest  # nr-bundle-spec / python
import json
from pathlib import Path

# Pull the full dataset (or use allow_patterns=["crp_<id>/*"] for one bundle).
local = snapshot_download(repo_id="NullRabbit/nr-bundles-public",
                          repo_type="dataset")

# Iterate bundles
for bundle_dir in sorted(Path(local).glob("crp_*")):
    manifest = BundleManifest.model_validate_json(
        (bundle_dir / "manifest.json").read_text()
    )
    print(manifest.primitive_id, manifest.family_id, manifest.fidelity_class)
    # Each modality is a Parquet file, read with pyarrow / pandas / polars
    # as the modality demands. See nr-bundle-spec for per-modality schemas.
    # Check manifest.files.<modality_parquet> before reading; bundles do
    # not contain modalities that were not captured during the workload.

datasets.load_dataset("NullRabbit/nr-bundles-public") does not work on this dataset: HF's auto-loader assumes a single tabular schema, which doesn't match the multi-modal bundle layout. Use snapshot_download + the bundle-spec parsers as shown above.

For schema, validation, and reference parsers (Python + Rust), see nr-bundle-spec. A validator CLI is available at tools/validate_bundle.py.

Licensing

• Dataset: CC-BY-4.0. Attribution required; commercial use permitted.
• Format spec: MIT (separate, at nr-bundle-spec).

The split is intentional. The format is freely reusable code/schema; the data carries attribution requirements so downstream work that uses NullRabbit's specific bundles can be traced.

Citation

@dataset{nullrabbit_nrbundlespublic_2026,
  author       = {NullRabbit},
  title        = {nr-bundles-public: a curated multi-modal bundle dataset for adversarial blockchain validator research},
  year         = {2026},
  month        = may,
  version      = {1},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/NullRabbit/nr-bundles-public},
  note         = {Format specified by nr-bundle-spec v0.2.0 (https://github.com/NullRabbitLabs/nr-bundle-spec).},
}

Related:

• Bundle format specification: nr-bundle-spec
• Earned autonomy paper (governance layer): Zenodo DOI 10.5281/zenodo.18406828
• Substrate paper (data-layer methodology): in preparation
• NullRabbit Labs: huggingface.co/NullRabbit
• Website: nullrabbit.ai

Contact

Research enquiries: security@nullrabbit.ai

Format extensions, schema feedback, or bundles produced by external researchers against the open spec, open an issue at nr-bundle-spec.