{"_canonicalization":{"envelope_id":"axm_ + sha256(envelope minus {signature, axiom_id, anchors})","envelope_signature":"ed25519(envelope minus {signature, axiom_id})","json":"sort_keys=True, separators=(',',':'), ensure_ascii=False, allow_nan=False, utf-8","leaf_hash":"sha256(0x00 || canonical_json(envelope_full))","seal_signature":"ed25519(seal minus {signature, sig_algorithm})"},"axiom_id":"axm_a3974e0b3f10df51d1921b4fcf8b628f861f4c9ea20329db450827e56ca64559","bitcoin_anchor":{"bitcoin_attestations":[],"calendar_attestations":[],"ots_url":"","stamped_at":"","status":"pending_next_stamp"},"envelope":{"anchors":[{"chain":"crovia.axiom_graph","height":0,"merkle_proof":"spider_vendor_press_v1","root_at_anchor":"spider_vendor_press_v1"}],"axiom_id":"axm_a3974e0b3f10df51d1921b4fcf8b628f861f4c9ea20329db450827e56ca64559","axiom_type":"AX.OBS","body":{"axiom_subtype":"news.vendor_press.v1","category":"news","fingerprint":"f30110e15d9d536a43dd262a19cf869269e02bd10e07dfcbf108871e111163e4","published":"Mon, 29 Jun 2026 00:00:00 -0400","receipt_hash":"f30110e15d9d536a43dd262a19cf869269e02bd10e07dfcbf108871e111163e4","schema":"spider.news.vendor_press.v1","spider":"vendor_press","spider_record":{"axiom_subtype":"news.vendor_press.v1","category":"news","decision_hint":"POSITIVE","envelope_target":"AX.OBS","fingerprint":"f30110e15d9d536a43dd262a19cf869269e02bd10e07dfcbf108871e111163e4","observed_at":"2026-06-29T04:44:03.414425Z","parent_run_hash":"36b5ab5c57ae76dc9e1a863501c4d38f172868cfae31b4ba5baf3caffaafb2c4","published":"Mon, 29 Jun 2026 00:00:00 -0400","runtime_version":"0.1.0","schema":"spider.news.vendor_press.v1","source_status":200,"source_url":"https://export.arxiv.org/rss/cs.AI","spider":"vendor_press","summary_excerpt":"arXiv:2606.28153v1 Announce Type: cross \nAbstract: Jailbreak attacks bypass LLM safety alignment, yet their mechanisms remain poorly understood. We provide evidence that attacks do not comprehensively eliminate safety features, but instead selectively suppress specific attention heads. We identify two functionally differentiated types: Adversarially Compromised Heads (ACHs) concentrated in early layers, which are suppressed under attacks, and Safety-Aligned Heads (SAHs) in mid-layers, which maintain robust activations even when attacks succeed. Ablation studies support the causal role of ACHs and the contribution of SAHs to robust activations: suppressing a small number of ACHs is sufficient to induce jailbreak-like behavior on normally refused inputs, while removing SAHs substantially weakens mid-layer safety activations. Token-level attribution further shows that ACH suppression is driven specifically by attack-template tokens, providing a mechanistic account of why attacks can bypas","title":"Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models","url":"https://arxiv.org/abs/2606.28153","vendor":"arxiv_cs_ai"},"summary":"arXiv:2606.28153v1 Announce Type: cross \nAbstract: Jailbreak attacks bypass LLM safety alignment, yet their mechanisms remain poorly understood. We provide evidence that attacks do not comprehensively eliminate safety features, but instead selectively suppress specific attention heads. We identify two functionally differentiated types: Adversarially Compromised Heads (ACHs) concentrated in early layers, which are suppressed under attacks, and Safety-Aligned Heads (SAHs) in mid-layers, which maintain robust activations even when attacks succeed. Ablation studies support the causal role of ACHs and the contribution of SAHs to robust activations: suppressing a small number of ACHs is sufficient to induce jailbreak-like behavior on normally refused inputs, while removing SAHs substantially weakens mid-layer safety activations. Token-level attribution further shows that ACH suppression is driven specifically by attack-template tokens, providing a mechanistic account of why attacks can bypas","title":"Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models","vendor":"arxiv_cs_ai"},"confidence":{"method":"deterministic"},"decision":"POSITIVE","issued_at":"2026-06-29T04:44:03Z","notes":"Spider vendor_press (news) news.vendor_press.v1","object":{"captured_by":"crovia.spider.vendor_press","primary_source_url":"https://arxiv.org/abs/2606.28153"},"predecessors":[],"schema":"crovia.axiom.v1","signature":"ed25519:7f413c034296a3274fdad3a570f7139156a995ac1b50389ca89899bdc1f53fd9d5e603132304cb4e84502f43fc64094540f38dd207810ce3b85b2d85fd77c800","signer":"crovia.substrate","subject":{"observed_at":"2026-06-29T04:44:03Z","source_collector":"spider:vendor_press","target_id":"https://arxiv.org/abs/2606.28153"},"tsa":{"authority":"crovia.substrate.bootstrap","rfc3161_token":"{\"kind\":\"crovia.bootstrap.tsa\",\"source_jsonl\":\"/opt/crovia/spider/data/news/vendor_press_v1.jsonl\",\"source_seal_merkle_root\":\"spider_vendor_press_v1\",\"upgrade_path\":\"Sessione H \\u2014 OpenTimestamps weekly anchor\"}"},"zk_mode":"clear","zk_proof":null},"ledger":{"leaf_hash":"d062d971f6f3d7929aac36aa58f0c952f3f1658e6f828ae0f9632622326edba4","leaf_index":261430,"ledger_path":"/opt/crovia/substrate/axiom_ledger.jsonl"},"merkle_proof":{"hash_alg":"sha256","leaf_prefix":"0x00","node_prefix":"0x01","odd_leaf_rule":"duplicate_last","path":[{"sibling":"c1e1e8a819f03f4c208245ceb6c8f284f14a2ff38dc8900971114025bf9ca3f8","side":"right"},{"sibling":"dfef6b40a5a9cc9df989e5c06ffa07d159a48349d24a421f3d68eb4b2e8da070","side":"left"},{"sibling":"be0b40c53956f3b3fa34b60c0625cea636d9239784a3f1b110aa1e2afbebaa12","side":"left"},{"sibling":"84c14f8e01284da67c9e5ab5aad1d4fed2482a09643bd1241f99d01a8d56166e","side":"right"},{"sibling":"2822a86953ba809f6211c399bc5537914e533fc8ed9db9d6a73638c2f06c5b41","side":"left"},{"sibling":"28f6d579aa317feaa020f8d687c6d588f190996b8a9f473dfbb0a5df14ba3293","side":"left"},{"sibling":"f598b90b37d7f9301fa045df6c69e9936b54d84361d68c339f86d13ecd08335e","side":"right"},{"sibling":"4b4bda5fa1fb23989b8f6f192c36dadc8266c378c69d42945af35c9d6ab81a10","side":"right"},{"sibling":"6d1df65d14253ed3a13c03e25aa9adcd9a091d8062a82e8586bcaee74013b90e","side":"left"},{"sibling":"9f9daa9d12e65b219f34c92aec45450536b79a42b8892050d66961432ae28ed1","side":"right"},{"sibling":"b5725d7b0807dc6da32d9788f20057fa8726be38d30a9ebdabc605ae92739122","side":"left"},{"sibling":"e321b2cac14cbe28f76ccb7938249a40ff60cd5d2128b5634be046ea10e984b8","side":"left"},{"sibling":"5900dc6c7d13855af9d0385baf1691ec386df33e450c422af1cabe0a36e40ad8","side":"left"},{"sibling":"ae636ddee98c71ab7a7dc55ddfab70c7f710a2b6abfdf7a8b5d16a4017d1c0d1","side":"left"},{"sibling":"f29798d8bb6aa9900eab878992d9ff0c53266debd87472f31ab26a6a3fb55880","side":"left"},{"sibling":"a04392fb9f2a3a620840e3b3fecd93d12e6d224e481c162389d8ae64e7194599","side":"left"},{"sibling":"c300cf0154c136afc09b1702a0be98f4ba5b6dc5cf57e8cc714ec1eaf4196eff","side":"left"},{"sibling":"d841ad93efda0869e5eb97678f348f03f5caab4353e05ff4bf18f47fb945b822","side":"left"}]},"schema":"crovia.axiom_proof.v1","seal":{"first_collector_run_id":"","first_receipt_hash":"","jsonl_path":"/opt/crovia/substrate/axiom_ledger.jsonl","key_id":"430895f101d38164","last_collector_run_id":"","last_receipt_hash":"","leaf_count":261662,"merkle_root":"aa8865c239aa2eb6c8aa7c6250f56b3cd5709854a8a07f6a29eb4ddd8802cb6f","public_key_hex":"cf742e26f75669dc673cb5c0786a1ae23ae8ca19c347317192ce40c28a7ff25c","run_id":"hourly_json_retrofit_20260629T053701Z","schema":"crovia.seal.v1","seal_family_version":"crovia-seal-family/1","seal_kind":"substrate_batch","sealed_at":"2026-06-29T05:38:02Z","sig_algorithm":"ed25519","signature":"476329233e82fb35fba2552ddc5d1d75b2bdd8513bbd281e9c40a0b8e475df374a62dcd8b456b0c5e8815984f5b4bf0983ae95d2cf4d7412ebb13a433b933c0a","signer_version":"1.1.0"},"trust_root":{"key_id":"430895f101d38164","public_key_hex":"cf742e26f75669dc673cb5c0786a1ae23ae8ca19c347317192ce40c28a7ff25c","signature_algorithm":"ed25519","url":"/registry/canon/TRUST_ROOT.md"},"verifier":{"spec":"/registry/canon/AXIOM_RECEIPT_v1.md","url":"/v/axm_a3974e0b3f10df51d1921b4fcf8b628f861f4c9ea20329db450827e56ca64559"}}