{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:QBBVIGVVWEZTRRY4PITJKK765S","short_pith_number":"pith:QBBVIGVV","schema_version":"1.0","canonical_sha256":"8043541ab5b13338c71c7a26952bfeec949637184746909fb8eebac47ec68ea6","source":{"kind":"arxiv","id":"2312.14302","version":2},"attestation_state":"computed","paper":{"title":"Exploiting Novel GPT-4 APIs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"Adam Gleave, Euan McLean, Kellin Pelrine, Micha{\\l} Zaj\\k{a}c, Mohammad Taufeeque","submitted_at":"2023-12-21T21:22:41Z","abstract_excerpt":"Language model attacks typically assume one of two extreme threat models: full white-box access to model weights, or black-box access limited to a text generation API. However, real-world APIs are often more flexible than just text generation: these APIs expose \"gray-box\" access leading to new threat vectors. To explore this, we red-team three new functionalities exposed in the GPT-4 APIs: fine-tuning, function calling and knowledge retrieval. We find that fine-tuning a model on as few as 15 harmful examples or 100 benign examples can remove core safeguards from GPT-4, enabling a range of harm"},"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":"2312.14302","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CR","submitted_at":"2023-12-21T21:22:41Z","cross_cats_sorted":["cs.AI","cs.CL","cs.LG"],"title_canon_sha256":"a43ded27f1541836e9eb70b29308870d2532072aa0c09a7a0770ad01afd66b6a","abstract_canon_sha256":"8f71cfed80622b3dc46ddf216ce7c1c45d9071661ef6da5bcd999c5933aae4a5"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:51:45.469409Z","signature_b64":"TuyEbzddJVOyzvrK7ZvdqZv2MTGR66JVH/zpAGngdLTMDNSxcOLOeLEdCODXxfp8PsX6n2VSGRdCPZ2yQ7+wDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"8043541ab5b13338c71c7a26952bfeec949637184746909fb8eebac47ec68ea6","last_reissued_at":"2026-07-05T08:51:45.468939Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:51:45.468939Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Exploiting Novel GPT-4 APIs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"Adam Gleave, Euan McLean, Kellin Pelrine, Micha{\\l} Zaj\\k{a}c, Mohammad Taufeeque","submitted_at":"2023-12-21T21:22:41Z","abstract_excerpt":"Language model attacks typically assume one of two extreme threat models: full white-box access to model weights, or black-box access limited to a text generation API. However, real-world APIs are often more flexible than just text generation: these APIs expose \"gray-box\" access leading to new threat vectors. To explore this, we red-team three new functionalities exposed in the GPT-4 APIs: fine-tuning, function calling and knowledge retrieval. We find that fine-tuning a model on as few as 15 harmful examples or 100 benign examples can remove core safeguards from GPT-4, enabling a range of harm"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2312.14302","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/2312.14302/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":"2312.14302","created_at":"2026-07-05T08:51:45.468996+00:00"},{"alias_kind":"arxiv_version","alias_value":"2312.14302v2","created_at":"2026-07-05T08:51:45.468996+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2312.14302","created_at":"2026-07-05T08:51:45.468996+00:00"},{"alias_kind":"pith_short_12","alias_value":"QBBVIGVVWEZT","created_at":"2026-07-05T08:51:45.468996+00:00"},{"alias_kind":"pith_short_16","alias_value":"QBBVIGVVWEZTRRY4","created_at":"2026-07-05T08:51:45.468996+00:00"},{"alias_kind":"pith_short_8","alias_value":"QBBVIGVV","created_at":"2026-07-05T08:51:45.468996+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.19147","citing_title":"Be Kind, Rewrite: Benign Projections via Rewriting Defend Against LLM Data Poisoning Attacks","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2410.09024","citing_title":"AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02574","citing_title":"Understanding the Effects of Safety Unalignment on Large Language Models","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2406.11717","citing_title":"Refusal in Language Models Is Mediated by a Single Direction","ref_index":171,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/QBBVIGVVWEZTRRY4PITJKK765S","json":"https://pith.science/pith/QBBVIGVVWEZTRRY4PITJKK765S.json","graph_json":"https://pith.science/api/pith-number/QBBVIGVVWEZTRRY4PITJKK765S/graph.json","events_json":"https://pith.science/api/pith-number/QBBVIGVVWEZTRRY4PITJKK765S/events.json","paper":"https://pith.science/paper/QBBVIGVV"},"agent_actions":{"view_html":"https://pith.science/pith/QBBVIGVVWEZTRRY4PITJKK765S","download_json":"https://pith.science/pith/QBBVIGVVWEZTRRY4PITJKK765S.json","view_paper":"https://pith.science/paper/QBBVIGVV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2312.14302&json=true","fetch_graph":"https://pith.science/api/pith-number/QBBVIGVVWEZTRRY4PITJKK765S/graph.json","fetch_events":"https://pith.science/api/pith-number/QBBVIGVVWEZTRRY4PITJKK765S/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/QBBVIGVVWEZTRRY4PITJKK765S/action/timestamp_anchor","attest_storage":"https://pith.science/pith/QBBVIGVVWEZTRRY4PITJKK765S/action/storage_attestation","attest_author":"https://pith.science/pith/QBBVIGVVWEZTRRY4PITJKK765S/action/author_attestation","sign_citation":"https://pith.science/pith/QBBVIGVVWEZTRRY4PITJKK765S/action/citation_signature","submit_replication":"https://pith.science/pith/QBBVIGVVWEZTRRY4PITJKK765S/action/replication_record"}},"created_at":"2026-07-05T08:51:45.468996+00:00","updated_at":"2026-07-05T08:51:45.468996+00:00"}