{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:KMDB4M4SW37Q7AAX3YGUB7Y3DM","short_pith_number":"pith:KMDB4M4S","schema_version":"1.0","canonical_sha256":"53061e3392b6ff0f8017de0d40ff1b1b202626d6741dcbfdd1a7bade0ad9a3b0","source":{"kind":"arxiv","id":"2408.15221","version":2},"attestation_state":"computed","paper":{"title":"LLM Defenses Are Not Robust to Multi-Turn Human Jailbreaks Yet","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.CR","cs.CY"],"primary_cat":"cs.LG","authors_text":"Cristina Menghini, Hugh Zhang, Ian Steneker, Nathaniel Li, Riley Goodside, Summer Yue, Willow Primack, Zifan Wang, Ziwen Han","submitted_at":"2024-08-27T17:33:30Z","abstract_excerpt":"Recent large language model (LLM) defenses have greatly improved models' ability to refuse harmful queries, even when adversarially attacked. However, LLM defenses are primarily evaluated against automated adversarial attacks in a single turn of conversation, an insufficient threat model for real-world malicious use. We demonstrate that multi-turn human jailbreaks uncover significant vulnerabilities, exceeding 70% attack success rate (ASR) on HarmBench against defenses that report single-digit ASRs with automated single-turn attacks. Human jailbreaks also reveal vulnerabilities in machine unle"},"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":"2408.15221","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-08-27T17:33:30Z","cross_cats_sorted":["cs.CL","cs.CR","cs.CY"],"title_canon_sha256":"f934fce144a82232edcee5b21c075bb368e8981af2deb0d44f69211000d19d7e","abstract_canon_sha256":"786b40a8a1552e7d8707930c2f1bc338f718d2186e85c8698236fcc5f122d21d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:02:41.944426Z","signature_b64":"AzsOP4XZ0o8n7Z3H4LA8SZctFj//4oZppWvz3qJAorEABuOqYaxAd6kj9B2iBjU5Hel6Bo956jLAMoOo+qnDDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"53061e3392b6ff0f8017de0d40ff1b1b202626d6741dcbfdd1a7bade0ad9a3b0","last_reissued_at":"2026-07-05T09:02:41.943898Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:02:41.943898Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"LLM Defenses Are Not Robust to Multi-Turn Human Jailbreaks Yet","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.CR","cs.CY"],"primary_cat":"cs.LG","authors_text":"Cristina Menghini, Hugh Zhang, Ian Steneker, Nathaniel Li, Riley Goodside, Summer Yue, Willow Primack, Zifan Wang, Ziwen Han","submitted_at":"2024-08-27T17:33:30Z","abstract_excerpt":"Recent large language model (LLM) defenses have greatly improved models' ability to refuse harmful queries, even when adversarially attacked. However, LLM defenses are primarily evaluated against automated adversarial attacks in a single turn of conversation, an insufficient threat model for real-world malicious use. We demonstrate that multi-turn human jailbreaks uncover significant vulnerabilities, exceeding 70% attack success rate (ASR) on HarmBench against defenses that report single-digit ASRs with automated single-turn attacks. Human jailbreaks also reveal vulnerabilities in machine unle"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2408.15221","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/2408.15221/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":"2408.15221","created_at":"2026-07-05T09:02:41.943953+00:00"},{"alias_kind":"arxiv_version","alias_value":"2408.15221v2","created_at":"2026-07-05T09:02:41.943953+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2408.15221","created_at":"2026-07-05T09:02:41.943953+00:00"},{"alias_kind":"pith_short_12","alias_value":"KMDB4M4SW37Q","created_at":"2026-07-05T09:02:41.943953+00:00"},{"alias_kind":"pith_short_16","alias_value":"KMDB4M4SW37Q7AAX","created_at":"2026-07-05T09:02:41.943953+00:00"},{"alias_kind":"pith_short_8","alias_value":"KMDB4M4S","created_at":"2026-07-05T09:02:41.943953+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":16,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.25013","citing_title":"Do Thinking Tokens Help with Safety?","ref_index":73,"is_internal_anchor":false},{"citing_arxiv_id":"2606.24388","citing_title":"PHANTOM: A Large-Scale Dataset of Multimodal Adversarial Attacks for Vision-Language Models","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2606.20408","citing_title":"NRT-Bench: Benchmarking Multi-Turn Red-Teaming of LLM Operator Agents in Safety-Critical Control Rooms","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2607.01277","citing_title":"Cognitive Firewall: A Proactive, Zero-Trust, Multi-Gate Framework for LLM Safety","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2606.14517","citing_title":"From Shield to Target: Denial-of-Service Attacks on LLM-Based Agent Guardrails","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07808","citing_title":"Where Instruction Hierarchy Breaks: Diagnosing and Repairing Failures in Reasoning Language Models","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2606.04141","citing_title":"Caught in the Act(ivation): Toward Pre-Output and Multi-Turn Detection of Credential Exfiltration by LLM Agents","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02041","citing_title":"SentGuard: Sentence-Level Streaming Guardrails for Large Language Models","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2605.27671","citing_title":"Evolving and Detecting Multi-Turn Deception using Geometric Signatures","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2511.02356","citing_title":"ASTRA: An Automated Framework for Strategy Discovery, Retrieval, and Evolution for Jailbreaking LLMs","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2410.09024","citing_title":"AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2604.27818","citing_title":"MASCing: Configurable Mixture-of-Experts Behavior via Activation Steering Masks","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10326","citing_title":"Jailbreaking the Matrix: Nullspace Steering for Controlled Model Subversion","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01687","citing_title":"MultiBreak: A Scalable and Diverse Multi-turn Jailbreak Benchmark for Evaluating LLM Safety","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2604.11309","citing_title":"The Salami Slicing Threat: Exploiting Cumulative Risks in LLM Systems","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2604.07727","citing_title":"TrajGuard: Streaming Hidden-state Trajectory Detection for Decoding-time Jailbreak Defense","ref_index":13,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/KMDB4M4SW37Q7AAX3YGUB7Y3DM","json":"https://pith.science/pith/KMDB4M4SW37Q7AAX3YGUB7Y3DM.json","graph_json":"https://pith.science/api/pith-number/KMDB4M4SW37Q7AAX3YGUB7Y3DM/graph.json","events_json":"https://pith.science/api/pith-number/KMDB4M4SW37Q7AAX3YGUB7Y3DM/events.json","paper":"https://pith.science/paper/KMDB4M4S"},"agent_actions":{"view_html":"https://pith.science/pith/KMDB4M4SW37Q7AAX3YGUB7Y3DM","download_json":"https://pith.science/pith/KMDB4M4SW37Q7AAX3YGUB7Y3DM.json","view_paper":"https://pith.science/paper/KMDB4M4S","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2408.15221&json=true","fetch_graph":"https://pith.science/api/pith-number/KMDB4M4SW37Q7AAX3YGUB7Y3DM/graph.json","fetch_events":"https://pith.science/api/pith-number/KMDB4M4SW37Q7AAX3YGUB7Y3DM/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/KMDB4M4SW37Q7AAX3YGUB7Y3DM/action/timestamp_anchor","attest_storage":"https://pith.science/pith/KMDB4M4SW37Q7AAX3YGUB7Y3DM/action/storage_attestation","attest_author":"https://pith.science/pith/KMDB4M4SW37Q7AAX3YGUB7Y3DM/action/author_attestation","sign_citation":"https://pith.science/pith/KMDB4M4SW37Q7AAX3YGUB7Y3DM/action/citation_signature","submit_replication":"https://pith.science/pith/KMDB4M4SW37Q7AAX3YGUB7Y3DM/action/replication_record"}},"created_at":"2026-07-05T09:02:41.943953+00:00","updated_at":"2026-07-05T09:02:41.943953+00:00"}