{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:NIRAS6TEQZUTBMLGGGHXQESDYQ","short_pith_number":"pith:NIRAS6TE","schema_version":"1.0","canonical_sha256":"6a22097a64866930b166318f781243c403d443e33b63d6a25221aae88f2f9d33","source":{"kind":"arxiv","id":"2402.09063","version":2},"attestation_state":"computed","paper":{"title":"Soft Prompt Threats: Attacking Safety Alignment and Unlearning in Open-Source LLMs through the Embedding Space","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"David Dobre, Gauthier Gidel, Leo Schwinn, Sophie Xhonneux, Stephan Gunnemann","submitted_at":"2024-02-14T10:20:03Z","abstract_excerpt":"Current research in adversarial robustness of LLMs focuses on discrete input manipulations in the natural language space, which can be directly transferred to closed-source models. However, this approach neglects the steady progression of open-source models. As open-source models advance in capability, ensuring their safety also becomes increasingly imperative. Yet, attacks tailored to open-source LLMs that exploit full model access remain largely unexplored. We address this research gap and propose the embedding space attack, which directly attacks the continuous embedding representation of i"},"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":"2402.09063","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-02-14T10:20:03Z","cross_cats_sorted":[],"title_canon_sha256":"1f63913efaa848bc296d596831641f53070ec49531fcb2982fca5555e6d35267","abstract_canon_sha256":"f58781ba1a823e5e41e357ca4ac928a8bd348a56533046e56523bfba8edd28f9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:49:43.257432Z","signature_b64":"fBJ4W2wcOhIXLJXHQVdSpv1dhJtDXWmtIo5pG4Se5KS+B/GhicaHApo+z5Ihacbpe/NTS03QxCOmV98QEKqXDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6a22097a64866930b166318f781243c403d443e33b63d6a25221aae88f2f9d33","last_reissued_at":"2026-07-05T10:49:43.256997Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:49:43.256997Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Soft Prompt Threats: Attacking Safety Alignment and Unlearning in Open-Source LLMs through the Embedding Space","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"David Dobre, Gauthier Gidel, Leo Schwinn, Sophie Xhonneux, Stephan Gunnemann","submitted_at":"2024-02-14T10:20:03Z","abstract_excerpt":"Current research in adversarial robustness of LLMs focuses on discrete input manipulations in the natural language space, which can be directly transferred to closed-source models. However, this approach neglects the steady progression of open-source models. As open-source models advance in capability, ensuring their safety also becomes increasingly imperative. Yet, attacks tailored to open-source LLMs that exploit full model access remain largely unexplored. We address this research gap and propose the embedding space attack, which directly attacks the continuous embedding representation of i"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2402.09063","kind":"arxiv","version":2},"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/2402.09063/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":"2402.09063","created_at":"2026-07-05T10:49:43.257063+00:00"},{"alias_kind":"arxiv_version","alias_value":"2402.09063v2","created_at":"2026-07-05T10:49:43.257063+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2402.09063","created_at":"2026-07-05T10:49:43.257063+00:00"},{"alias_kind":"pith_short_12","alias_value":"NIRAS6TEQZUT","created_at":"2026-07-05T10:49:43.257063+00:00"},{"alias_kind":"pith_short_16","alias_value":"NIRAS6TEQZUTBMLG","created_at":"2026-07-05T10:49:43.257063+00:00"},{"alias_kind":"pith_short_8","alias_value":"NIRAS6TE","created_at":"2026-07-05T10:49:43.257063+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2503.02574","citing_title":"LLM-Safety Evaluations Lack Robustness","ref_index":49,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11170","citing_title":"Unlearning with Asymmetric Sources: Improved Unlearning-Utility Trade-off with Public Data","ref_index":129,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10403","citing_title":"Latent Instruction Representation Alignment: defending against jailbreaks, backdoors and undesired knowledge in LLMs","ref_index":32,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/NIRAS6TEQZUTBMLGGGHXQESDYQ","json":"https://pith.science/pith/NIRAS6TEQZUTBMLGGGHXQESDYQ.json","graph_json":"https://pith.science/api/pith-number/NIRAS6TEQZUTBMLGGGHXQESDYQ/graph.json","events_json":"https://pith.science/api/pith-number/NIRAS6TEQZUTBMLGGGHXQESDYQ/events.json","paper":"https://pith.science/paper/NIRAS6TE"},"agent_actions":{"view_html":"https://pith.science/pith/NIRAS6TEQZUTBMLGGGHXQESDYQ","download_json":"https://pith.science/pith/NIRAS6TEQZUTBMLGGGHXQESDYQ.json","view_paper":"https://pith.science/paper/NIRAS6TE","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2402.09063&json=true","fetch_graph":"https://pith.science/api/pith-number/NIRAS6TEQZUTBMLGGGHXQESDYQ/graph.json","fetch_events":"https://pith.science/api/pith-number/NIRAS6TEQZUTBMLGGGHXQESDYQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/NIRAS6TEQZUTBMLGGGHXQESDYQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/NIRAS6TEQZUTBMLGGGHXQESDYQ/action/storage_attestation","attest_author":"https://pith.science/pith/NIRAS6TEQZUTBMLGGGHXQESDYQ/action/author_attestation","sign_citation":"https://pith.science/pith/NIRAS6TEQZUTBMLGGGHXQESDYQ/action/citation_signature","submit_replication":"https://pith.science/pith/NIRAS6TEQZUTBMLGGGHXQESDYQ/action/replication_record"}},"created_at":"2026-07-05T10:49:43.257063+00:00","updated_at":"2026-07-05T10:49:43.257063+00:00"}