{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:PCCYBRRMUKWRX7MVNLKVN4FDGT","short_pith_number":"pith:PCCYBRRM","schema_version":"1.0","canonical_sha256":"788580c62ca2ad1bfd956ad556f0a334d77b9820a6062ac01b9751979fb44baf","source":{"kind":"arxiv","id":"2605.29901","version":1},"attestation_state":"computed","paper":{"title":"Dissecting the Black Box: Circuit-Level Analysis of LLM Vulnerability Detection","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CR","authors_text":"Christian Gehrmann, Chun Zhou, Syafiq Al Atiiq","submitted_at":"2026-05-28T13:23:13Z","abstract_excerpt":"Large language models (LLMs) can detect software vulnerabilities, but how do they actually identify vulnerable code? We address this question using mechanistic interpretability; analyzing the internal computations of a neural network to understand its reasoning process.Using Circuit Tracer on Gemma-2-2b, we trace the computational pathways activated when the model classifies 472 C/C++ code samples as vulnerable or safe. Our analysis reveals a surprising finding: the model primarily relies on safety detectors, attention heads that recognize safe coding patterns, rather than directly detecting v"},"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":"2605.29901","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CR","submitted_at":"2026-05-28T13:23:13Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"b9eae35f7b9e826af90aa5f3674e71b7904024a87aeecc29694b82bd97e903ee","abstract_canon_sha256":"256998ba9810f39ef939ee8d14b46d919136ca34ad33ecc7bf74e2f4b6ee33bf"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-29T02:05:58.654063Z","signature_b64":"Kl8ccXhAijPX6AomftuXUcqvLGjOKibtB7cofwWVNqB8c+BmO1mJbRI8LgoUY4b7p3OBWEt1LoYTdy+Cx/MxDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"788580c62ca2ad1bfd956ad556f0a334d77b9820a6062ac01b9751979fb44baf","last_reissued_at":"2026-05-29T02:05:58.653304Z","signature_status":"signed_v1","first_computed_at":"2026-05-29T02:05:58.653304Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Dissecting the Black Box: Circuit-Level Analysis of LLM Vulnerability Detection","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CR","authors_text":"Christian Gehrmann, Chun Zhou, Syafiq Al Atiiq","submitted_at":"2026-05-28T13:23:13Z","abstract_excerpt":"Large language models (LLMs) can detect software vulnerabilities, but how do they actually identify vulnerable code? We address this question using mechanistic interpretability; analyzing the internal computations of a neural network to understand its reasoning process.Using Circuit Tracer on Gemma-2-2b, we trace the computational pathways activated when the model classifies 472 C/C++ code samples as vulnerable or safe. Our analysis reveals a surprising finding: the model primarily relies on safety detectors, attention heads that recognize safe coding patterns, rather than directly detecting v"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.29901","kind":"arxiv","version":1},"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/2605.29901/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":"2605.29901","created_at":"2026-05-29T02:05:58.653447+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.29901v1","created_at":"2026-05-29T02:05:58.653447+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.29901","created_at":"2026-05-29T02:05:58.653447+00:00"},{"alias_kind":"pith_short_12","alias_value":"PCCYBRRMUKWR","created_at":"2026-05-29T02:05:58.653447+00:00"},{"alias_kind":"pith_short_16","alias_value":"PCCYBRRMUKWRX7MV","created_at":"2026-05-29T02:05:58.653447+00:00"},{"alias_kind":"pith_short_8","alias_value":"PCCYBRRM","created_at":"2026-05-29T02:05:58.653447+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PCCYBRRMUKWRX7MVNLKVN4FDGT","json":"https://pith.science/pith/PCCYBRRMUKWRX7MVNLKVN4FDGT.json","graph_json":"https://pith.science/api/pith-number/PCCYBRRMUKWRX7MVNLKVN4FDGT/graph.json","events_json":"https://pith.science/api/pith-number/PCCYBRRMUKWRX7MVNLKVN4FDGT/events.json","paper":"https://pith.science/paper/PCCYBRRM"},"agent_actions":{"view_html":"https://pith.science/pith/PCCYBRRMUKWRX7MVNLKVN4FDGT","download_json":"https://pith.science/pith/PCCYBRRMUKWRX7MVNLKVN4FDGT.json","view_paper":"https://pith.science/paper/PCCYBRRM","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.29901&json=true","fetch_graph":"https://pith.science/api/pith-number/PCCYBRRMUKWRX7MVNLKVN4FDGT/graph.json","fetch_events":"https://pith.science/api/pith-number/PCCYBRRMUKWRX7MVNLKVN4FDGT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PCCYBRRMUKWRX7MVNLKVN4FDGT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PCCYBRRMUKWRX7MVNLKVN4FDGT/action/storage_attestation","attest_author":"https://pith.science/pith/PCCYBRRMUKWRX7MVNLKVN4FDGT/action/author_attestation","sign_citation":"https://pith.science/pith/PCCYBRRMUKWRX7MVNLKVN4FDGT/action/citation_signature","submit_replication":"https://pith.science/pith/PCCYBRRMUKWRX7MVNLKVN4FDGT/action/replication_record"}},"created_at":"2026-05-29T02:05:58.653447+00:00","updated_at":"2026-05-29T02:05:58.653447+00:00"}