{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:NWMCIAQQX5TZ63VWJEHOD6CRRO","short_pith_number":"pith:NWMCIAQQ","schema_version":"1.0","canonical_sha256":"6d98240210bf679f6eb6490ee1f8518baa7528477756da1dfea5ce2ce081ebe9","source":{"kind":"arxiv","id":"2602.13379","version":2},"attestation_state":"computed","paper":{"title":"Unsafer in Many Turns: Benchmarking and Defending Multi-Turn Safety Risks in Tool-Using Agents","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG","cs.SE"],"primary_cat":"cs.CR","authors_text":"Bo Li, Dawn Song, Minzhou Pan, Simon Yu, Weiyan Shi, Xue Lin, Xu Li, Yiyou Sun","submitted_at":"2026-02-13T18:38:18Z","abstract_excerpt":"LLM-based agents are becoming increasingly capable, yet their safety lags behind. This creates a gap between what agents can do and should do. This gap widens as agents engage in multi-turn interactions and employ diverse tools, introducing new risks overlooked by existing benchmarks. To systematically scale safety testing into multi-turn, tool-realistic settings, we propose a principled taxonomy that transforms single-turn harmful tasks into multi-turn attack sequences. Using this taxonomy, we construct MT-AgentRisk (Multi-Turn Agent Risk Benchmark), the first benchmark to evaluate multi-turn"},"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":"2602.13379","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.CR","submitted_at":"2026-02-13T18:38:18Z","cross_cats_sorted":["cs.AI","cs.CL","cs.LG","cs.SE"],"title_canon_sha256":"aac9a2b7a6cc1c7c0bb919def4b518c639cc06088b59ca415dd6b7dffbf9d1e0","abstract_canon_sha256":"c37f5c14e9f42e31d8d269f3fac0c54cf428f24bcf8d0418e9e565b49210da13"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-12T01:08:22.797107Z","signature_b64":"0YVaxsNXCstfEgEKuYCBM9kULrSsOm8d3/P8BOsI9cRLTDif1vaapZY3jMOYex7vSaTzAY4yWr1IeMV0CNuiCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6d98240210bf679f6eb6490ee1f8518baa7528477756da1dfea5ce2ce081ebe9","last_reissued_at":"2026-06-12T01:08:22.796038Z","signature_status":"signed_v1","first_computed_at":"2026-06-12T01:08:22.796038Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Unsafer in Many Turns: Benchmarking and Defending Multi-Turn Safety Risks in Tool-Using Agents","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG","cs.SE"],"primary_cat":"cs.CR","authors_text":"Bo Li, Dawn Song, Minzhou Pan, Simon Yu, Weiyan Shi, Xue Lin, Xu Li, Yiyou Sun","submitted_at":"2026-02-13T18:38:18Z","abstract_excerpt":"LLM-based agents are becoming increasingly capable, yet their safety lags behind. This creates a gap between what agents can do and should do. This gap widens as agents engage in multi-turn interactions and employ diverse tools, introducing new risks overlooked by existing benchmarks. To systematically scale safety testing into multi-turn, tool-realistic settings, we propose a principled taxonomy that transforms single-turn harmful tasks into multi-turn attack sequences. Using this taxonomy, we construct MT-AgentRisk (Multi-Turn Agent Risk Benchmark), the first benchmark to evaluate multi-turn"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2602.13379","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/2602.13379/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":"2602.13379","created_at":"2026-06-12T01:08:22.796193+00:00"},{"alias_kind":"arxiv_version","alias_value":"2602.13379v2","created_at":"2026-06-12T01:08:22.796193+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.13379","created_at":"2026-06-12T01:08:22.796193+00:00"},{"alias_kind":"pith_short_12","alias_value":"NWMCIAQQX5TZ","created_at":"2026-06-12T01:08:22.796193+00:00"},{"alias_kind":"pith_short_16","alias_value":"NWMCIAQQX5TZ63VW","created_at":"2026-06-12T01:08:22.796193+00:00"},{"alias_kind":"pith_short_8","alias_value":"NWMCIAQQ","created_at":"2026-06-12T01:08:22.796193+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":4,"sample":[{"citing_arxiv_id":"2605.22643","citing_title":"Boiling the Frog: A Multi-Turn Benchmark for Agentic Safety","ref_index":48,"is_internal_anchor":true},{"citing_arxiv_id":"2605.22643","citing_title":"Boiling the Frog: A Multi-Turn Benchmark for Agentic Safety","ref_index":48,"is_internal_anchor":true},{"citing_arxiv_id":"2605.17453","citing_title":"Trust No Tool: Evaluating and Defending LLM Agents under Untrusted Tool Feedback","ref_index":33,"is_internal_anchor":true},{"citing_arxiv_id":"2603.09002","citing_title":"Security Considerations for Multi-agent Systems","ref_index":118,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/NWMCIAQQX5TZ63VWJEHOD6CRRO","json":"https://pith.science/pith/NWMCIAQQX5TZ63VWJEHOD6CRRO.json","graph_json":"https://pith.science/api/pith-number/NWMCIAQQX5TZ63VWJEHOD6CRRO/graph.json","events_json":"https://pith.science/api/pith-number/NWMCIAQQX5TZ63VWJEHOD6CRRO/events.json","paper":"https://pith.science/paper/NWMCIAQQ"},"agent_actions":{"view_html":"https://pith.science/pith/NWMCIAQQX5TZ63VWJEHOD6CRRO","download_json":"https://pith.science/pith/NWMCIAQQX5TZ63VWJEHOD6CRRO.json","view_paper":"https://pith.science/paper/NWMCIAQQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2602.13379&json=true","fetch_graph":"https://pith.science/api/pith-number/NWMCIAQQX5TZ63VWJEHOD6CRRO/graph.json","fetch_events":"https://pith.science/api/pith-number/NWMCIAQQX5TZ63VWJEHOD6CRRO/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/NWMCIAQQX5TZ63VWJEHOD6CRRO/action/timestamp_anchor","attest_storage":"https://pith.science/pith/NWMCIAQQX5TZ63VWJEHOD6CRRO/action/storage_attestation","attest_author":"https://pith.science/pith/NWMCIAQQX5TZ63VWJEHOD6CRRO/action/author_attestation","sign_citation":"https://pith.science/pith/NWMCIAQQX5TZ63VWJEHOD6CRRO/action/citation_signature","submit_replication":"https://pith.science/pith/NWMCIAQQX5TZ63VWJEHOD6CRRO/action/replication_record"}},"created_at":"2026-06-12T01:08:22.796193+00:00","updated_at":"2026-06-12T01:08:22.796193+00:00"}