{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:DFSYMWYD6ZAYYXVOZ5WDOVHIBT","short_pith_number":"pith:DFSYMWYD","schema_version":"1.0","canonical_sha256":"1965865b03f6418c5eaecf6c3754e80cf0fa98401f717999072549bb0a3c7021","source":{"kind":"arxiv","id":"2606.27091","version":1},"attestation_state":"computed","paper":{"title":"Inherited Circuits, Learned Semantics: How Fine-Tuning Creates Evasion Vulnerabilities Invisible to Standard Evaluation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CR","authors_text":"Ryan Fetterman","submitted_at":"2026-06-25T14:29:03Z","abstract_excerpt":"LLMs fine-tuned for security classification are usually evaluated on held-out examples from the same distribution as their training data. We show that this can miss vulnerabilities introduced by fine-tuning itself: models can learn token-level indicator semantics that preserve canonical accuracy while failing under behavior-preserving transformations such as PowerShell alias substitution, command reconstruction, string construction, execution indirection, and case mutation. We study Foundation-Sec-8B-Instruct and its base model, Llama-3.1-8B-Instruct, on matched PowerShell classification cohor"},"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":"2606.27091","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CR","submitted_at":"2026-06-25T14:29:03Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"deadf20c934033d0c96099ed9d772592251091e2a6028410a6eee5c7146ee3c0","abstract_canon_sha256":"32ba45eb297dc7f97f6fe211e38bbdf87353d973ea90062c448a5ed3e59bfc62"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-26T01:16:09.190786Z","signature_b64":"wCLMS5Q/Uesj2Ju1yt+Dp+lBEon+oaNxw55muf1G7tF7TF/diDVmCNahwugRAttIKRjbVlYZOXWtDL/UYn/zBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1965865b03f6418c5eaecf6c3754e80cf0fa98401f717999072549bb0a3c7021","last_reissued_at":"2026-06-26T01:16:09.190415Z","signature_status":"signed_v1","first_computed_at":"2026-06-26T01:16:09.190415Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Inherited Circuits, Learned Semantics: How Fine-Tuning Creates Evasion Vulnerabilities Invisible to Standard Evaluation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CR","authors_text":"Ryan Fetterman","submitted_at":"2026-06-25T14:29:03Z","abstract_excerpt":"LLMs fine-tuned for security classification are usually evaluated on held-out examples from the same distribution as their training data. We show that this can miss vulnerabilities introduced by fine-tuning itself: models can learn token-level indicator semantics that preserve canonical accuracy while failing under behavior-preserving transformations such as PowerShell alias substitution, command reconstruction, string construction, execution indirection, and case mutation. We study Foundation-Sec-8B-Instruct and its base model, Llama-3.1-8B-Instruct, on matched PowerShell classification cohor"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.27091","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/2606.27091/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":"2606.27091","created_at":"2026-06-26T01:16:09.190470+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.27091v1","created_at":"2026-06-26T01:16:09.190470+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.27091","created_at":"2026-06-26T01:16:09.190470+00:00"},{"alias_kind":"pith_short_12","alias_value":"DFSYMWYD6ZAY","created_at":"2026-06-26T01:16:09.190470+00:00"},{"alias_kind":"pith_short_16","alias_value":"DFSYMWYD6ZAYYXVO","created_at":"2026-06-26T01:16:09.190470+00:00"},{"alias_kind":"pith_short_8","alias_value":"DFSYMWYD","created_at":"2026-06-26T01:16:09.190470+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/DFSYMWYD6ZAYYXVOZ5WDOVHIBT","json":"https://pith.science/pith/DFSYMWYD6ZAYYXVOZ5WDOVHIBT.json","graph_json":"https://pith.science/api/pith-number/DFSYMWYD6ZAYYXVOZ5WDOVHIBT/graph.json","events_json":"https://pith.science/api/pith-number/DFSYMWYD6ZAYYXVOZ5WDOVHIBT/events.json","paper":"https://pith.science/paper/DFSYMWYD"},"agent_actions":{"view_html":"https://pith.science/pith/DFSYMWYD6ZAYYXVOZ5WDOVHIBT","download_json":"https://pith.science/pith/DFSYMWYD6ZAYYXVOZ5WDOVHIBT.json","view_paper":"https://pith.science/paper/DFSYMWYD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.27091&json=true","fetch_graph":"https://pith.science/api/pith-number/DFSYMWYD6ZAYYXVOZ5WDOVHIBT/graph.json","fetch_events":"https://pith.science/api/pith-number/DFSYMWYD6ZAYYXVOZ5WDOVHIBT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/DFSYMWYD6ZAYYXVOZ5WDOVHIBT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/DFSYMWYD6ZAYYXVOZ5WDOVHIBT/action/storage_attestation","attest_author":"https://pith.science/pith/DFSYMWYD6ZAYYXVOZ5WDOVHIBT/action/author_attestation","sign_citation":"https://pith.science/pith/DFSYMWYD6ZAYYXVOZ5WDOVHIBT/action/citation_signature","submit_replication":"https://pith.science/pith/DFSYMWYD6ZAYYXVOZ5WDOVHIBT/action/replication_record"}},"created_at":"2026-06-26T01:16:09.190470+00:00","updated_at":"2026-06-26T01:16:09.190470+00:00"}