{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:CQ445C746WIGW6BDTQXYHMWICT","short_pith_number":"pith:CQ445C74","schema_version":"1.0","canonical_sha256":"1439ce8bfcf5906b78239c2f83b2c814f8dc44852a974542099c38c797f0764a","source":{"kind":"arxiv","id":"2605.21706","version":1},"attestation_state":"computed","paper":{"title":"Latent-space Attacks for Refusal Evasion in Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Battista Biggio, Fabio Brau, Fabio Roli, Giorgio Piras, Luca Oneto, Maura Pintor, Raffaele Mura","submitted_at":"2026-05-20T20:10:27Z","abstract_excerpt":"Safety-aligned language models are trained to refuse harmful requests, yet refusal behavior can be suppressed by steering their internal representations. Existing methods do so by ablating a refusal direction from model activations, aiming to remove refusal from the model's residual stream. Despite their empirical success, these methods lack a principled account of the latent-space transformation they induce and why it suppresses refusal. In this work, we recast refusal suppression as a latent-space evasion attack against linear probes trained to separate refused from answered prompts. Under t"},"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":"2605.21706","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-20T20:10:27Z","cross_cats_sorted":[],"title_canon_sha256":"5ffe54f8167887b5380ff2946232931878e0c01e7bb655df66cb3d79f3040ca0","abstract_canon_sha256":"b9ddfd7c9e4f5e1dd085afb5343cece6294dc555a026caf2de8fd407b6fd3589"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-22T01:03:28.799856Z","signature_b64":"Lxlr0UDqBJqWWaUBj53Vt2G9C9AHMEvwp1czh+qAanN6a7ylUhTmvpQ493ub2TGNbt2ZuwAoYBrNpvT97fibAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1439ce8bfcf5906b78239c2f83b2c814f8dc44852a974542099c38c797f0764a","last_reissued_at":"2026-05-22T01:03:28.799371Z","signature_status":"signed_v1","first_computed_at":"2026-05-22T01:03:28.799371Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Latent-space Attacks for Refusal Evasion in Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Battista Biggio, Fabio Brau, Fabio Roli, Giorgio Piras, Luca Oneto, Maura Pintor, Raffaele Mura","submitted_at":"2026-05-20T20:10:27Z","abstract_excerpt":"Safety-aligned language models are trained to refuse harmful requests, yet refusal behavior can be suppressed by steering their internal representations. Existing methods do so by ablating a refusal direction from model activations, aiming to remove refusal from the model's residual stream. Despite their empirical success, these methods lack a principled account of the latent-space transformation they induce and why it suppresses refusal. In this work, we recast refusal suppression as a latent-space evasion attack against linear probes trained to separate refused from answered prompts. Under t"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.21706","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/2605.21706/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":"2605.21706","created_at":"2026-05-22T01:03:28.799452+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.21706v1","created_at":"2026-05-22T01:03:28.799452+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.21706","created_at":"2026-05-22T01:03:28.799452+00:00"},{"alias_kind":"pith_short_12","alias_value":"CQ445C746WIG","created_at":"2026-05-22T01:03:28.799452+00:00"},{"alias_kind":"pith_short_16","alias_value":"CQ445C746WIGW6BD","created_at":"2026-05-22T01:03:28.799452+00:00"},{"alias_kind":"pith_short_8","alias_value":"CQ445C74","created_at":"2026-05-22T01:03:28.799452+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/CQ445C746WIGW6BDTQXYHMWICT","json":"https://pith.science/pith/CQ445C746WIGW6BDTQXYHMWICT.json","graph_json":"https://pith.science/api/pith-number/CQ445C746WIGW6BDTQXYHMWICT/graph.json","events_json":"https://pith.science/api/pith-number/CQ445C746WIGW6BDTQXYHMWICT/events.json","paper":"https://pith.science/paper/CQ445C74"},"agent_actions":{"view_html":"https://pith.science/pith/CQ445C746WIGW6BDTQXYHMWICT","download_json":"https://pith.science/pith/CQ445C746WIGW6BDTQXYHMWICT.json","view_paper":"https://pith.science/paper/CQ445C74","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.21706&json=true","fetch_graph":"https://pith.science/api/pith-number/CQ445C746WIGW6BDTQXYHMWICT/graph.json","fetch_events":"https://pith.science/api/pith-number/CQ445C746WIGW6BDTQXYHMWICT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CQ445C746WIGW6BDTQXYHMWICT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CQ445C746WIGW6BDTQXYHMWICT/action/storage_attestation","attest_author":"https://pith.science/pith/CQ445C746WIGW6BDTQXYHMWICT/action/author_attestation","sign_citation":"https://pith.science/pith/CQ445C746WIGW6BDTQXYHMWICT/action/citation_signature","submit_replication":"https://pith.science/pith/CQ445C746WIGW6BDTQXYHMWICT/action/replication_record"}},"created_at":"2026-05-22T01:03:28.799452+00:00","updated_at":"2026-05-22T01:03:28.799452+00:00"}