{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:2T5IQMSX4MYJJU2DXQ4W55YZFD","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"2a3f685c21e3536239d4e7eb3c3674fabea9b16ff3ba2c2182c8d5a13c2d0909","cross_cats_sorted":["cs.AI","cs.CR"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-04T23:30:29Z","title_canon_sha256":"8960708f235e99c37ce82d273ec3d0dfc3d5230cb674c314ecd322d93ff6ea0f"},"schema_version":"1.0","source":{"id":"2605.03226","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.03226","created_at":"2026-06-09T02:08:43Z"},{"alias_kind":"arxiv_version","alias_value":"2605.03226v2","created_at":"2026-06-09T02:08:43Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.03226","created_at":"2026-06-09T02:08:43Z"},{"alias_kind":"pith_short_12","alias_value":"2T5IQMSX4MYJ","created_at":"2026-06-09T02:08:43Z"},{"alias_kind":"pith_short_16","alias_value":"2T5IQMSX4MYJJU2D","created_at":"2026-06-09T02:08:43Z"},{"alias_kind":"pith_short_8","alias_value":"2T5IQMSX","created_at":"2026-06-09T02:08:43Z"}],"graph_snapshots":[{"event_id":"sha256:cb789604b9cda68533ae059b811d61ff184dcf0de2ac0874f65ddca2b092d0b6","target":"graph","created_at":"2026-06-09T02:08:43Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"On Llama-3-8B-Instruct and Llama-3.2-3B-Instruct, this approach cuts the WildJailbreak attack success rate from 11.5% and 20.1% down to 1-3%, but pushes refusal on jailbreak-shaped benign prompts from 14-22% to 74-94%. Interleaving the same hard prompts 1:1 with adversarially-framed benign prompts cuts that refusal back down to 30-51% on 8B and 52-72% on 3B, at a cost of 2-6 percentage points of attack success rate."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the frequency with which the model's own rollouts are judged harmful provides a reliable and unbiased measure of prompt difficulty for safety fine-tuning, and that the external harm judgment itself is consistent and free of systematic error."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Self-mined hardness from model rollouts reduces WildJailbreak attack success rates to 1-3% on Llama models but increases over-refusal on benign prompts, which mixing with adversarially-framed benign prompts partially mitigates."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Models can improve safety fine-tuning by selecting prompts based on how often their own responses are judged harmful."}],"snapshot_sha256":"3d7e31b15bb10f819dab1db20e5f3ebbcba4f29c74aeb78ea690417b533169d4"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"cff818b5e7fc17a1ca754d37c5d18dd7efd6104dc5d1512a4b9080fdd2460396"},"integrity":{"available":true,"clean":true,"detectors_run":[{"findings_count":0,"name":"ai_meta_artifact","ran_at":"2026-05-20T14:35:20.977131Z","status":"completed","version":"1.0.0"},{"findings_count":0,"name":"doi_title_agreement","ran_at":"2026-05-20T01:31:21.804826Z","status":"completed","version":"1.0.0"},{"findings_count":0,"name":"doi_compliance","ran_at":"2026-05-19T15:34:04.492328Z","status":"completed","version":"1.0.0"}],"endpoint":"/pith/2605.03226/integrity.json","findings":[],"snapshot_sha256":"63b72441310a29bbde61c85889cfa895137a0ec66dbff06e0cbab195626790f7","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Safety fine-tuning of language models typically requires a curated adversarial dataset. We take a different approach: score each candidate prompt's difficulty by how often the target model's own rollouts are judged harmful, then fine-tune on the hardest prompts paired with the model's own non-jailbroken rollouts. On Llama-3-8B-Instruct and Llama-3.2-3B-Instruct, this approach cuts the WildJailbreak attack success rate from 11.5% and 20.1% down to 1-3%, but pushes refusal on jailbreak-shaped benign prompts from 14-22% to 74-94%. Interleaving the same hard prompts 1:1 with adversarially-framed b","authors_text":"Donghua Zhang, Garv Shah, Prakhar Gupta","cross_cats":["cs.AI","cs.CR"],"headline":"Models can improve safety fine-tuning by selecting prompts based on how often their own responses are judged harmful.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-04T23:30:29Z","title":"Self-Mined Hardness for Safety Fine-Tuning"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.03226","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-08T18:17:53.452364Z","id":"8e991f83-c42a-4de8-8aef-93c71675d900","model_set":{"reader":"grok-4.3"},"one_line_summary":"Self-mined hardness from model rollouts reduces WildJailbreak attack success rates to 1-3% on Llama models but increases over-refusal on benign prompts, which mixing with adversarially-framed benign prompts partially mitigates.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Models can improve safety fine-tuning by selecting prompts based on how often their own responses are judged harmful.","strongest_claim":"On Llama-3-8B-Instruct and Llama-3.2-3B-Instruct, this approach cuts the WildJailbreak attack success rate from 11.5% and 20.1% down to 1-3%, but pushes refusal on jailbreak-shaped benign prompts from 14-22% to 74-94%. Interleaving the same hard prompts 1:1 with adversarially-framed benign prompts cuts that refusal back down to 30-51% on 8B and 52-72% on 3B, at a cost of 2-6 percentage points of attack success rate.","weakest_assumption":"That the frequency with which the model's own rollouts are judged harmful provides a reliable and unbiased measure of prompt difficulty for safety fine-tuning, and that the external harm judgment itself is consistent and free of systematic error."}},"verdict_id":"8e991f83-c42a-4de8-8aef-93c71675d900"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:5091d1e0949803706a58c675e8da8ba48ece793192181a7868ede2c8bad80b9c","target":"record","created_at":"2026-06-09T02:08:43Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"2a3f685c21e3536239d4e7eb3c3674fabea9b16ff3ba2c2182c8d5a13c2d0909","cross_cats_sorted":["cs.AI","cs.CR"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-04T23:30:29Z","title_canon_sha256":"8960708f235e99c37ce82d273ec3d0dfc3d5230cb674c314ecd322d93ff6ea0f"},"schema_version":"1.0","source":{"id":"2605.03226","kind":"arxiv","version":2}},"canonical_sha256":"d4fa883257e33094d343bc396ef71928d8522a248c903bea260f58104eb70458","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"d4fa883257e33094d343bc396ef71928d8522a248c903bea260f58104eb70458","first_computed_at":"2026-06-09T02:08:43.559849Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-06-09T02:08:43.559849Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"OlfKrvfBpTKdWwQ/ZbHvYWPUF/qRiDhUqfIq9wH+7IiIqsTH0nUvAqlnvxDzW+gXVkV38oYNT+matTbr3+nrBg==","signature_status":"signed_v1","signed_at":"2026-06-09T02:08:43.560701Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.03226","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:5091d1e0949803706a58c675e8da8ba48ece793192181a7868ede2c8bad80b9c","sha256:cb789604b9cda68533ae059b811d61ff184dcf0de2ac0874f65ddca2b092d0b6"],"state_sha256":"de95d540d2f804e5e281966c9785770376a77c356a50f3e8244e05aa9d680854"}