{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:IA77M4ON7NIPNGJQLWMUVZIOF5","short_pith_number":"pith:IA77M4ON","schema_version":"1.0","canonical_sha256":"403ff671cdfb50f699305d994ae50e2f73f7e0cc96e241b08575684fe3094748","source":{"kind":"arxiv","id":"2508.03153","version":2},"attestation_state":"computed","paper":{"title":"Estimating Worst-Case Frontier Risks of Open-Weight LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Chris Koch, Eric Wallace, Kai Chen, Miles Wang, Olivia Watkins","submitted_at":"2025-08-05T06:57:53Z","abstract_excerpt":"In this paper, we study the worst-case frontier risks of releasing gpt-oss. We introduce malicious fine-tuning (MFT), where we attempt to elicit maximum capabilities by fine-tuning gpt-oss to be as capable as possible in two domains: biology and cybersecurity. To maximize biological risk (biorisk), we curate tasks related to threat creation and train gpt-oss in an RL environment with web browsing. To maximize cybersecurity risk, we train gpt-oss in an agentic coding environment to solve capture-the-flag (CTF) challenges. We compare these MFT models against open- and closed-weight LLMs on front"},"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":"2508.03153","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-08-05T06:57:53Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"28d1053d51c763027a032e4839bdcf0abc130d265a583c67772d96c7102b9df7","abstract_canon_sha256":"aab593bbd4cb81fc755f6e61eb50efeacb271a53428028d7b8b44039295b90de"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:53:06.789656Z","signature_b64":"Q5IIg8lRoypvJNMvkrCjHuNuVRCOO/lxciX9ZQn/TcQADJLsI8mrsbz+zzvuzonv+vKtqRSfq5xHrBwVFvjaAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"403ff671cdfb50f699305d994ae50e2f73f7e0cc96e241b08575684fe3094748","last_reissued_at":"2026-07-05T11:53:06.789138Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:53:06.789138Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Estimating Worst-Case Frontier Risks of Open-Weight LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Chris Koch, Eric Wallace, Kai Chen, Miles Wang, Olivia Watkins","submitted_at":"2025-08-05T06:57:53Z","abstract_excerpt":"In this paper, we study the worst-case frontier risks of releasing gpt-oss. We introduce malicious fine-tuning (MFT), where we attempt to elicit maximum capabilities by fine-tuning gpt-oss to be as capable as possible in two domains: biology and cybersecurity. To maximize biological risk (biorisk), we curate tasks related to threat creation and train gpt-oss in an RL environment with web browsing. To maximize cybersecurity risk, we train gpt-oss in an agentic coding environment to solve capture-the-flag (CTF) challenges. We compare these MFT models against open- and closed-weight LLMs on front"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2508.03153","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/2508.03153/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":"2508.03153","created_at":"2026-07-05T11:53:06.789202+00:00"},{"alias_kind":"arxiv_version","alias_value":"2508.03153v2","created_at":"2026-07-05T11:53:06.789202+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2508.03153","created_at":"2026-07-05T11:53:06.789202+00:00"},{"alias_kind":"pith_short_12","alias_value":"IA77M4ON7NIP","created_at":"2026-07-05T11:53:06.789202+00:00"},{"alias_kind":"pith_short_16","alias_value":"IA77M4ON7NIPNGJQ","created_at":"2026-07-05T11:53:06.789202+00:00"},{"alias_kind":"pith_short_8","alias_value":"IA77M4ON","created_at":"2026-07-05T11:53:06.789202+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.26526","citing_title":"Open-Weight LLM Fine-Tuning Defenses are Susceptible to Simple Attacks","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2508.20697","citing_title":"Token Buncher: Shielding LLMs from Harmful Reinforcement Learning Fine-Tuning","ref_index":49,"is_internal_anchor":false},{"citing_arxiv_id":"2604.22167","citing_title":"Estimating Tail Risks in Language Model Output Distributions","ref_index":41,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/IA77M4ON7NIPNGJQLWMUVZIOF5","json":"https://pith.science/pith/IA77M4ON7NIPNGJQLWMUVZIOF5.json","graph_json":"https://pith.science/api/pith-number/IA77M4ON7NIPNGJQLWMUVZIOF5/graph.json","events_json":"https://pith.science/api/pith-number/IA77M4ON7NIPNGJQLWMUVZIOF5/events.json","paper":"https://pith.science/paper/IA77M4ON"},"agent_actions":{"view_html":"https://pith.science/pith/IA77M4ON7NIPNGJQLWMUVZIOF5","download_json":"https://pith.science/pith/IA77M4ON7NIPNGJQLWMUVZIOF5.json","view_paper":"https://pith.science/paper/IA77M4ON","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2508.03153&json=true","fetch_graph":"https://pith.science/api/pith-number/IA77M4ON7NIPNGJQLWMUVZIOF5/graph.json","fetch_events":"https://pith.science/api/pith-number/IA77M4ON7NIPNGJQLWMUVZIOF5/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/IA77M4ON7NIPNGJQLWMUVZIOF5/action/timestamp_anchor","attest_storage":"https://pith.science/pith/IA77M4ON7NIPNGJQLWMUVZIOF5/action/storage_attestation","attest_author":"https://pith.science/pith/IA77M4ON7NIPNGJQLWMUVZIOF5/action/author_attestation","sign_citation":"https://pith.science/pith/IA77M4ON7NIPNGJQLWMUVZIOF5/action/citation_signature","submit_replication":"https://pith.science/pith/IA77M4ON7NIPNGJQLWMUVZIOF5/action/replication_record"}},"created_at":"2026-07-05T11:53:06.789202+00:00","updated_at":"2026-07-05T11:53:06.789202+00:00"}