{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:OP6TCEI4PW4TO65MHAGKAJ5YL2","short_pith_number":"pith:OP6TCEI4","schema_version":"1.0","canonical_sha256":"73fd31111c7db9377bac380ca027b85eba5c72cd6c8ee3a617f140aded3f7dcb","source":{"kind":"arxiv","id":"2606.05958","version":1},"attestation_state":"computed","paper":{"title":"Steering Vectors are an Adversarial Attack Surface","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Abzal Aidakhmetov, Adrian Robert Minut, Donato Crisostomi, Emanuele Rodol\\`a, Iacopo Masi, Tommaso Mencattini","submitted_at":"2026-06-04T09:56:48Z","abstract_excerpt":"Activation steering has become a popular way to control Large Language Model (LLM) behavior without fine-tuning. Since the technique is plug-and-play, users share datasets and precomputed vectors to steer model activations. However, we show that a \\emph{stealth data poisoning attack} silently compromises this pipeline. By substituting $4{-}6\\%$ of tokens in the steering dataset, an attacker can silently align the resulting vector with an anti-refusal direction. This jailbreaks the target model while preserving the intended steering effect on benign prompts. Under this threat model, a malicious"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2606.05958","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-06-04T09:56:48Z","cross_cats_sorted":[],"title_canon_sha256":"24013bbb337d35c973333dd1200fab8577f1b860814cb8b6094b52f73d86587c","abstract_canon_sha256":"456f5e08530b49a35be497ad427bd2b778adafb0f3a10d5c825150c9850b6456"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-05T01:15:28.795581Z","signature_b64":"lMwf3sWxCwRfp4kuE72NG9jU5bwc1PNLKnBKuD9/52Pq7FJvIYMQlpRVCwuyofS+IqprT7BAVqMocyn01GqtAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"73fd31111c7db9377bac380ca027b85eba5c72cd6c8ee3a617f140aded3f7dcb","last_reissued_at":"2026-06-05T01:15:28.795125Z","signature_status":"signed_v1","first_computed_at":"2026-06-05T01:15:28.795125Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Steering Vectors are an Adversarial Attack Surface","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Abzal Aidakhmetov, Adrian Robert Minut, Donato Crisostomi, Emanuele Rodol\\`a, Iacopo Masi, Tommaso Mencattini","submitted_at":"2026-06-04T09:56:48Z","abstract_excerpt":"Activation steering has become a popular way to control Large Language Model (LLM) behavior without fine-tuning. Since the technique is plug-and-play, users share datasets and precomputed vectors to steer model activations. However, we show that a \\emph{stealth data poisoning attack} silently compromises this pipeline. By substituting $4{-}6\\%$ of tokens in the steering dataset, an attacker can silently align the resulting vector with an anti-refusal direction. This jailbreaks the target model while preserving the intended steering effect on benign prompts. Under this threat model, a malicious"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.05958","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2606.05958/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2606.05958","created_at":"2026-06-05T01:15:28.795181+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.05958v1","created_at":"2026-06-05T01:15:28.795181+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.05958","created_at":"2026-06-05T01:15:28.795181+00:00"},{"alias_kind":"pith_short_12","alias_value":"OP6TCEI4PW4T","created_at":"2026-06-05T01:15:28.795181+00:00"},{"alias_kind":"pith_short_16","alias_value":"OP6TCEI4PW4TO65M","created_at":"2026-06-05T01:15:28.795181+00:00"},{"alias_kind":"pith_short_8","alias_value":"OP6TCEI4","created_at":"2026-06-05T01:15:28.795181+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OP6TCEI4PW4TO65MHAGKAJ5YL2","json":"https://pith.science/pith/OP6TCEI4PW4TO65MHAGKAJ5YL2.json","graph_json":"https://pith.science/api/pith-number/OP6TCEI4PW4TO65MHAGKAJ5YL2/graph.json","events_json":"https://pith.science/api/pith-number/OP6TCEI4PW4TO65MHAGKAJ5YL2/events.json","paper":"https://pith.science/paper/OP6TCEI4"},"agent_actions":{"view_html":"https://pith.science/pith/OP6TCEI4PW4TO65MHAGKAJ5YL2","download_json":"https://pith.science/pith/OP6TCEI4PW4TO65MHAGKAJ5YL2.json","view_paper":"https://pith.science/paper/OP6TCEI4","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.05958&json=true","fetch_graph":"https://pith.science/api/pith-number/OP6TCEI4PW4TO65MHAGKAJ5YL2/graph.json","fetch_events":"https://pith.science/api/pith-number/OP6TCEI4PW4TO65MHAGKAJ5YL2/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OP6TCEI4PW4TO65MHAGKAJ5YL2/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OP6TCEI4PW4TO65MHAGKAJ5YL2/action/storage_attestation","attest_author":"https://pith.science/pith/OP6TCEI4PW4TO65MHAGKAJ5YL2/action/author_attestation","sign_citation":"https://pith.science/pith/OP6TCEI4PW4TO65MHAGKAJ5YL2/action/citation_signature","submit_replication":"https://pith.science/pith/OP6TCEI4PW4TO65MHAGKAJ5YL2/action/replication_record"}},"created_at":"2026-06-05T01:15:28.795181+00:00","updated_at":"2026-06-05T01:15:28.795181+00:00"}