{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:G6GX5GMYJN44JMQXCCDUFXPQZO","short_pith_number":"pith:G6GX5GMY","schema_version":"1.0","canonical_sha256":"378d7e99984b79c4b217108742ddf0cb9c419bf4817f2d85dae524fe0aac6e46","source":{"kind":"arxiv","id":"2501.07927","version":3},"attestation_state":"computed","paper":{"title":"Gandalf the Red: Adaptive Security for LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.CR"],"primary_cat":"cs.LG","authors_text":"Adri\\`a Romero-L\\'opez, Alan Commike, Andreas Signer, Athanasios Theocharis, Dami\\'an Pascual-Ortiz, Daniel Timbrell, David Haber, Jakub Podolak, Janet Darling, Julia Bazi\\'nska, Kyriacos Shiarlis, Manuel Knott, Marco Lancini, Mateo Rojas-Carulla, Matthew Fiedler, Matthias Kraft, Max Mathys, Mykhailo Bichurin, Niklas Pfister, Peter Dienes, Samuel Trautwein, Samuel Watts, Santiago Arias, V\\'aclav Volhejn, Yun-Han Wu, Zsolt Terek","submitted_at":"2025-01-14T08:30:49Z","abstract_excerpt":"Current evaluations of defenses against prompt attacks in large language model (LLM) applications often overlook two critical factors: the dynamic nature of adversarial behavior and the usability penalties imposed on legitimate users by restrictive defenses. We propose D-SEC (Dynamic Security Utility Threat Model), which explicitly separates attackers from legitimate users, models multi-step interactions, and expresses the security-utility in an optimizable form. We further address the shortcomings in existing evaluations by introducing Gandalf, a crowd-sourced, gamified red-teaming platform d"},"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":"2501.07927","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-01-14T08:30:49Z","cross_cats_sorted":["cs.AI","cs.CL","cs.CR"],"title_canon_sha256":"8d24f446c7eab8814a3fc30fe08a4ae27d2ddfa02e77e79624cb83e1806bd7a2","abstract_canon_sha256":"a2e31253c669a4f7610ef8ffaf06cb0cb8b911fa3ac43002e823adb57fc1885d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:47:59.646591Z","signature_b64":"Gvs7fqAqB8SdShkUy1VtjptMVfz+/+j1MP0Y3WBbXB0W6PJHQEHmlyutCLTW7KSAonk6Pcrn2VKg1CHL+4J6CQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"378d7e99984b79c4b217108742ddf0cb9c419bf4817f2d85dae524fe0aac6e46","last_reissued_at":"2026-07-05T11:47:59.646076Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:47:59.646076Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Gandalf the Red: Adaptive Security for LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.CR"],"primary_cat":"cs.LG","authors_text":"Adri\\`a Romero-L\\'opez, Alan Commike, Andreas Signer, Athanasios Theocharis, Dami\\'an Pascual-Ortiz, Daniel Timbrell, David Haber, Jakub Podolak, Janet Darling, Julia Bazi\\'nska, Kyriacos Shiarlis, Manuel Knott, Marco Lancini, Mateo Rojas-Carulla, Matthew Fiedler, Matthias Kraft, Max Mathys, Mykhailo Bichurin, Niklas Pfister, Peter Dienes, Samuel Trautwein, Samuel Watts, Santiago Arias, V\\'aclav Volhejn, Yun-Han Wu, Zsolt Terek","submitted_at":"2025-01-14T08:30:49Z","abstract_excerpt":"Current evaluations of defenses against prompt attacks in large language model (LLM) applications often overlook two critical factors: the dynamic nature of adversarial behavior and the usability penalties imposed on legitimate users by restrictive defenses. We propose D-SEC (Dynamic Security Utility Threat Model), which explicitly separates attackers from legitimate users, models multi-step interactions, and expresses the security-utility in an optimizable form. We further address the shortcomings in existing evaluations by introducing Gandalf, a crowd-sourced, gamified red-teaming platform d"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2501.07927","kind":"arxiv","version":3},"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/2501.07927/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":"2501.07927","created_at":"2026-07-05T11:47:59.646137+00:00"},{"alias_kind":"arxiv_version","alias_value":"2501.07927v3","created_at":"2026-07-05T11:47:59.646137+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2501.07927","created_at":"2026-07-05T11:47:59.646137+00:00"},{"alias_kind":"pith_short_12","alias_value":"G6GX5GMYJN44","created_at":"2026-07-05T11:47:59.646137+00:00"},{"alias_kind":"pith_short_16","alias_value":"G6GX5GMYJN44JMQX","created_at":"2026-07-05T11:47:59.646137+00:00"},{"alias_kind":"pith_short_8","alias_value":"G6GX5GMY","created_at":"2026-07-05T11:47:59.646137+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2506.13726","citing_title":"Weakest Link in the Chain: Security Vulnerabilities in Advanced Reasoning Models","ref_index":15,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/G6GX5GMYJN44JMQXCCDUFXPQZO","json":"https://pith.science/pith/G6GX5GMYJN44JMQXCCDUFXPQZO.json","graph_json":"https://pith.science/api/pith-number/G6GX5GMYJN44JMQXCCDUFXPQZO/graph.json","events_json":"https://pith.science/api/pith-number/G6GX5GMYJN44JMQXCCDUFXPQZO/events.json","paper":"https://pith.science/paper/G6GX5GMY"},"agent_actions":{"view_html":"https://pith.science/pith/G6GX5GMYJN44JMQXCCDUFXPQZO","download_json":"https://pith.science/pith/G6GX5GMYJN44JMQXCCDUFXPQZO.json","view_paper":"https://pith.science/paper/G6GX5GMY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2501.07927&json=true","fetch_graph":"https://pith.science/api/pith-number/G6GX5GMYJN44JMQXCCDUFXPQZO/graph.json","fetch_events":"https://pith.science/api/pith-number/G6GX5GMYJN44JMQXCCDUFXPQZO/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/G6GX5GMYJN44JMQXCCDUFXPQZO/action/timestamp_anchor","attest_storage":"https://pith.science/pith/G6GX5GMYJN44JMQXCCDUFXPQZO/action/storage_attestation","attest_author":"https://pith.science/pith/G6GX5GMYJN44JMQXCCDUFXPQZO/action/author_attestation","sign_citation":"https://pith.science/pith/G6GX5GMYJN44JMQXCCDUFXPQZO/action/citation_signature","submit_replication":"https://pith.science/pith/G6GX5GMYJN44JMQXCCDUFXPQZO/action/replication_record"}},"created_at":"2026-07-05T11:47:59.646137+00:00","updated_at":"2026-07-05T11:47:59.646137+00:00"}