{"_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_5e9b2eca50a3373dc5c7cb59202ebfbc8508c9369c413fb3e586e68c21da336b","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_5e9b2eca50a3373dc5c7cb59202ebfbc8508c9369c413fb3e586e68c21da336b","axiom_type":"AX.OBS","body":{"axiom_subtype":"news.vendor_press.v1","category":"news","fingerprint":"0d3c7e15752d0dcb4d99afc54f245c957180bd9ede9f088728737c1cee01e89f","published":"Thu, 18 Jun 2026 00:00:00 -0400","receipt_hash":"0d3c7e15752d0dcb4d99afc54f245c957180bd9ede9f088728737c1cee01e89f","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":"0d3c7e15752d0dcb4d99afc54f245c957180bd9ede9f088728737c1cee01e89f","observed_at":"2026-06-18T04:43:37.219665Z","parent_run_hash":"de79a40f7b3537d88842f7ac355e799c5df2adcb4fc32a4e28096d4bbdf01739","published":"Thu, 18 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:2604.23130v2 Announce Type: replace-cross \nAbstract: Jailbreak attacks expose a persistent failure mode in safety-aligned LLMs: models can be pushed into harmful behavior, but the internal representations enabling this shift remain poorly localized. Recent mechanistic safety studies often explain such behavior through broad representational objects, including global refusal directions, activation steering vectors, and refusal-related SAE features. We instead ask whether jailbreak vulnerability can be traced to finer-grained, prompt-conditioned SAE feature subgroups. We introduce a token-driven mechanistic pipeline that decomposes the residual stream of Gemma-2-2B into Sparse Autoencoder (SAE) features and identifies feature subgroups associated with unsafe behavior. Using single-category unsafe examples from BeaverTails to reduce cross-category interference, we extract harmful concepts from adversarial responses and align them with concept-relevant prompt tokens through subspace ","title":"From Concept-Aligned Tokens to Vulnerable Features: Mechanistic Localization of Jailbreaks","url":"https://arxiv.org/abs/2604.23130","vendor":"arxiv_cs_ai"},"summary":"arXiv:2604.23130v2 Announce Type: replace-cross \nAbstract: Jailbreak attacks expose a persistent failure mode in safety-aligned LLMs: models can be pushed into harmful behavior, but the internal representations enabling this shift remain poorly localized. Recent mechanistic safety studies often explain such behavior through broad representational objects, including global refusal directions, activation steering vectors, and refusal-related SAE features. We instead ask whether jailbreak vulnerability can be traced to finer-grained, prompt-conditioned SAE feature subgroups. We introduce a token-driven mechanistic pipeline that decomposes the residual stream of Gemma-2-2B into Sparse Autoencoder (SAE) features and identifies feature subgroups associated with unsafe behavior. Using single-category unsafe examples from BeaverTails to reduce cross-category interference, we extract harmful concepts from adversarial responses and align them with concept-relevant prompt tokens through subspace ","title":"From Concept-Aligned Tokens to Vulnerable Features: Mechanistic Localization of Jailbreaks","vendor":"arxiv_cs_ai"},"confidence":{"method":"deterministic"},"decision":"POSITIVE","issued_at":"2026-06-18T04:43:37Z","notes":"Spider vendor_press (news) news.vendor_press.v1","object":{"captured_by":"crovia.spider.vendor_press","primary_source_url":"https://arxiv.org/abs/2604.23130"},"predecessors":[],"schema":"crovia.axiom.v1","signature":"ed25519:0902438ab8a9ec735db33e63d2ed27a01d9417d5d03a02941f4976897280d3a7dff3a0a87f733c6ccf7fac526725e01cc936dd51ab1e603c80e6217696475108","signer":"crovia.substrate","subject":{"observed_at":"2026-06-18T04:43:37Z","source_collector":"spider:vendor_press","target_id":"https://arxiv.org/abs/2604.23130"},"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":"af5e32c688c9721be7daba138706e62628f5f5f4ae6eaff75700f0cd8ca6f67a","leaf_index":233505,"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":"24a4f43e04c2fa669f9c0212207fd2685486a9a5ebec4aede6af8408e326ce83","side":"left"},{"sibling":"ee88b52f299a4ede9c5978a09311370ad5da5c6a653f0fc3236889b9a56adf4d","side":"right"},{"sibling":"68704eef93147ecc98e195e9096438520d8369ad83bb5438548f4fab60f2b384","side":"right"},{"sibling":"1ba02a06c10a51cf25d99ca1d85c63d9fcc0392b5ab9b376c64fb24a6a50938a","side":"right"},{"sibling":"ccef570cc5460ba72a01b2a0b3b6e7eee19dfe0476c303523c1d0b6d5e6c19ba","side":"right"},{"sibling":"c42f991fcbafbe53d8fc33e1c5c2d1d51ef70e0c2709828f03f40cba8d7b8389","side":"left"},{"sibling":"5e436eec5370aaf4cccb6c432bd004f0a554eab0e19ec87f70ba66470ad7303b","side":"right"},{"sibling":"fe9c643bdeb268143f61f15d89d0ff9dd03becb2914656c7a873d95f7266b5ce","side":"right"},{"sibling":"ad9e1cf26277141407aebd8692bfb16135dd1e614913e54b1dd445fed15d105a","side":"right"},{"sibling":"b151db1a7de0ce9a329250fae8b690f5b55ab3cfe5468a1fbb5f5bde0703b420","side":"right"},{"sibling":"7dc9143c057343b46a3b988492fba5262dec443cc1fb6070e3ea81543ca6a522","side":"right"},{"sibling":"ad5850946feb9a22361b5b9df0884f9ef1edcc7efea0374ff5782080ccb1a947","side":"right"},{"sibling":"410c633928fea11c5b4bdddb431956b1d7c320db9cda00d2fe32e0fcf888d7b7","side":"left"},{"sibling":"e616c34dbaf9456d5a6d3e2da82cde8621293c9f6d8a4cf9e441d7fd9cc81579","side":"right"},{"sibling":"94c0c932e61657f5e37fdba43f6ca9eddea8359425a7c1558dabe566911d5304","side":"right"},{"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":234491,"merkle_root":"02576a6980e38bab47864ae2c57b5a5ff21e554e9bdf8f64bdf28155ff1aabec","public_key_hex":"cf742e26f75669dc673cb5c0786a1ae23ae8ca19c347317192ce40c28a7ff25c","run_id":"hourly_json_retrofit_20260618T143732Z","schema":"crovia.seal.v1","seal_family_version":"crovia-seal-family/1","seal_kind":"substrate_batch","sealed_at":"2026-06-18T18:33:39Z","sig_algorithm":"ed25519","signature":"b6c708778fc38b7789a2b91156cfe87252a7cd3a1d29121cba11a0c78f8cf104ca3019fc50a962fa5a216bcc4922fc8f3f69c04d1f8332c6dc0d931e1012e502","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_5e9b2eca50a3373dc5c7cb59202ebfbc8508c9369c413fb3e586e68c21da336b"}}