{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:S2EXWZWRVNQRZBLRPZYMYRDX54","short_pith_number":"pith:S2EXWZWR","schema_version":"1.0","canonical_sha256":"96897b66d1ab611c85717e70cc4477ef2f2e563dcdfcda6e14e27a992c89c326","source":{"kind":"arxiv","id":"2410.22770","version":3},"attestation_state":"computed","paper":{"title":"InjecGuard: Benchmarking and Mitigating Over-defense in Prompt Injection Guardrail Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.CL","authors_text":"Hao Li, Xiaogeng Liu","submitted_at":"2024-10-30T07:39:42Z","abstract_excerpt":"Prompt injection attacks pose a critical threat to large language models (LLMs), enabling goal hijacking and data leakage. Prompt guard models, though effective in defense, suffer from over-defense -- falsely flagging benign inputs as malicious due to trigger word bias. To address this issue, we introduce NotInject, an evaluation dataset that systematically measures over-defense across various prompt guard models. NotInject contains 339 benign samples enriched with trigger words common in prompt injection attacks, enabling fine-grained evaluation. Our results show that state-of-the-art models "},"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":"2410.22770","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-10-30T07:39:42Z","cross_cats_sorted":["cs.AI","cs.CR"],"title_canon_sha256":"0a95a57701417c57eeb5c2b988b6d772c72ebb93ca89f5713e6f557cb2af5af6","abstract_canon_sha256":"ea16bb990772f2095d0a91e82d468b21ad42d68b3a62e378ec250ac531337849"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:41:16.766904Z","signature_b64":"0dcbCnBhrZ3gPLJkOAt4vMUKhIXNQBPUiN/YQrSnilBNizkE/ekkxbIMbG3YZwN0uqqXuZ2S4ejdQJrQ7/PmCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"96897b66d1ab611c85717e70cc4477ef2f2e563dcdfcda6e14e27a992c89c326","last_reissued_at":"2026-07-05T10:41:16.766321Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:41:16.766321Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"InjecGuard: Benchmarking and Mitigating Over-defense in Prompt Injection Guardrail Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.CL","authors_text":"Hao Li, Xiaogeng Liu","submitted_at":"2024-10-30T07:39:42Z","abstract_excerpt":"Prompt injection attacks pose a critical threat to large language models (LLMs), enabling goal hijacking and data leakage. Prompt guard models, though effective in defense, suffer from over-defense -- falsely flagging benign inputs as malicious due to trigger word bias. To address this issue, we introduce NotInject, an evaluation dataset that systematically measures over-defense across various prompt guard models. NotInject contains 339 benign samples enriched with trigger words common in prompt injection attacks, enabling fine-grained evaluation. Our results show that state-of-the-art models "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.22770","kind":"arxiv","version":3},"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/2410.22770/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":"2410.22770","created_at":"2026-07-05T10:41:16.766452+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.22770v3","created_at":"2026-07-05T10:41:16.766452+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.22770","created_at":"2026-07-05T10:41:16.766452+00:00"},{"alias_kind":"pith_short_12","alias_value":"S2EXWZWRVNQR","created_at":"2026-07-05T10:41:16.766452+00:00"},{"alias_kind":"pith_short_16","alias_value":"S2EXWZWRVNQRZBLR","created_at":"2026-07-05T10:41:16.766452+00:00"},{"alias_kind":"pith_short_8","alias_value":"S2EXWZWR","created_at":"2026-07-05T10:41:16.766452+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":13,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.26377","citing_title":"Verifying Intent and Harm: A Unified Defense Against LLM-Generated Threats","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2606.22659","citing_title":"Confidently Wrong: Severity-Aware Calibration of Prompt-Injection Detectors under Attack Shift","ref_index":6,"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":7,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02959","citing_title":"Gate AI: LLM Security Benchmark Evaluation Methodology and Results","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2605.26999","citing_title":"Prompt Injection Detection is Regime-Dependent: A Deployment-Aware Evaluation with Interpretable Structural Signals","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11268","citing_title":"Context-Aware Spear Phishing: Generative AI-Enabled Attacks Against Individuals via Public Social Media Data","ref_index":65,"is_internal_anchor":false},{"citing_arxiv_id":"2605.03378","citing_title":"ARGUS: Defending LLM Agents Against Context-Aware Prompt Injection","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2604.25716","citing_title":"Cross-Lingual Jailbreak Detection via Semantic Codebooks","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2604.25109","citing_title":"Structured Security Auditing and Robustness Enhancement for Untrusted Agent Skills","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2604.21477","citing_title":"MCP Pitfall Lab: Exposing Developer Pitfalls in MCP Tool Server Security under Multi-Vector Attacks","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2604.07536","citing_title":"TRUSTDESC: Preventing Tool Poisoning in LLM Applications via Trusted Description Generation","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2604.16543","citing_title":"Conjunctive Prompt Attacks in Multi-Agent LLM Systems","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2604.17562","citing_title":"SafeAgent: A Runtime Protection Architecture for Agentic Systems","ref_index":13,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/S2EXWZWRVNQRZBLRPZYMYRDX54","json":"https://pith.science/pith/S2EXWZWRVNQRZBLRPZYMYRDX54.json","graph_json":"https://pith.science/api/pith-number/S2EXWZWRVNQRZBLRPZYMYRDX54/graph.json","events_json":"https://pith.science/api/pith-number/S2EXWZWRVNQRZBLRPZYMYRDX54/events.json","paper":"https://pith.science/paper/S2EXWZWR"},"agent_actions":{"view_html":"https://pith.science/pith/S2EXWZWRVNQRZBLRPZYMYRDX54","download_json":"https://pith.science/pith/S2EXWZWRVNQRZBLRPZYMYRDX54.json","view_paper":"https://pith.science/paper/S2EXWZWR","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.22770&json=true","fetch_graph":"https://pith.science/api/pith-number/S2EXWZWRVNQRZBLRPZYMYRDX54/graph.json","fetch_events":"https://pith.science/api/pith-number/S2EXWZWRVNQRZBLRPZYMYRDX54/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/S2EXWZWRVNQRZBLRPZYMYRDX54/action/timestamp_anchor","attest_storage":"https://pith.science/pith/S2EXWZWRVNQRZBLRPZYMYRDX54/action/storage_attestation","attest_author":"https://pith.science/pith/S2EXWZWRVNQRZBLRPZYMYRDX54/action/author_attestation","sign_citation":"https://pith.science/pith/S2EXWZWRVNQRZBLRPZYMYRDX54/action/citation_signature","submit_replication":"https://pith.science/pith/S2EXWZWRVNQRZBLRPZYMYRDX54/action/replication_record"}},"created_at":"2026-07-05T10:41:16.766452+00:00","updated_at":"2026-07-05T10:41:16.766452+00:00"}