{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:YEJU657KWRV3SKWAOLTA2XSEFL","short_pith_number":"pith:YEJU657K","schema_version":"1.0","canonical_sha256":"c1134f77eab46bb92ac072e60d5e442ac72989b3cef0330587816b751743d116","source":{"kind":"arxiv","id":"2404.05993","version":2},"attestation_state":"computed","paper":{"title":"AEGIS: Online Adaptive AI Content Safety Moderation with Ensemble of LLM Experts","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.CY"],"primary_cat":"cs.LG","authors_text":"Christopher Parisien, Erick Galinkin, Prasoon Varshney, Shaona Ghosh","submitted_at":"2024-04-09T03:54:28Z","abstract_excerpt":"As Large Language Models (LLMs) and generative AI become more widespread, the content safety risks associated with their use also increase. We find a notable deficiency in high-quality content safety datasets and benchmarks that comprehensively cover a wide range of critical safety areas. To address this, we define a broad content safety risk taxonomy, comprising 13 critical risk and 9 sparse risk categories. Additionally, we curate AEGISSAFETYDATASET, a new dataset of approximately 26, 000 human-LLM interaction instances, complete with human annotations adhering to the taxonomy. We plan to re"},"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":"2404.05993","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-04-09T03:54:28Z","cross_cats_sorted":["cs.CL","cs.CY"],"title_canon_sha256":"ae4e9fc05795b1444c7f8be80e95de4028df2b136a8f3efb925d35bc11c3abf0","abstract_canon_sha256":"2d7e7725cc0153ffa1ba7c898b32447a98b1dd41b25f77ad44ba9a91255a6c69"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:05:54.292232Z","signature_b64":"nIx0z+Lj62RPi9LtA1uS60tR8ub5hHmBHH2Jc3W7M9lxV9WU5lI5vI6OP1xVu4fRqbBjHO6+0vzw4LxclqZtCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"c1134f77eab46bb92ac072e60d5e442ac72989b3cef0330587816b751743d116","last_reissued_at":"2026-07-05T09:05:54.291821Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:05:54.291821Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"AEGIS: Online Adaptive AI Content Safety Moderation with Ensemble of LLM Experts","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.CY"],"primary_cat":"cs.LG","authors_text":"Christopher Parisien, Erick Galinkin, Prasoon Varshney, Shaona Ghosh","submitted_at":"2024-04-09T03:54:28Z","abstract_excerpt":"As Large Language Models (LLMs) and generative AI become more widespread, the content safety risks associated with their use also increase. We find a notable deficiency in high-quality content safety datasets and benchmarks that comprehensively cover a wide range of critical safety areas. To address this, we define a broad content safety risk taxonomy, comprising 13 critical risk and 9 sparse risk categories. Additionally, we curate AEGISSAFETYDATASET, a new dataset of approximately 26, 000 human-LLM interaction instances, complete with human annotations adhering to the taxonomy. We plan to re"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2404.05993","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/2404.05993/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":"2404.05993","created_at":"2026-07-05T09:05:54.291870+00:00"},{"alias_kind":"arxiv_version","alias_value":"2404.05993v2","created_at":"2026-07-05T09:05:54.291870+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.05993","created_at":"2026-07-05T09:05:54.291870+00:00"},{"alias_kind":"pith_short_12","alias_value":"YEJU657KWRV3","created_at":"2026-07-05T09:05:54.291870+00:00"},{"alias_kind":"pith_short_16","alias_value":"YEJU657KWRV3SKWA","created_at":"2026-07-05T09:05:54.291870+00:00"},{"alias_kind":"pith_short_8","alias_value":"YEJU657K","created_at":"2026-07-05T09:05:54.291870+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":23,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.06326","citing_title":"DT-Guard: Intent-Driven Reasoning-Active Training for Reasoning-Free LLM Safety Guardrail","ref_index":12,"is_internal_anchor":true},{"citing_arxiv_id":"2606.25034","citing_title":"Yuvion VL: A Multimodal Foundation Model for Adversarial Content and AI Safety","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08044","citing_title":"When Behavioral Safety Evaluation Fails: A Representation-Level Perspective","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05936","citing_title":"Epistemic Injustice in Language Models: An Audit of Pretraining Filters and Guardrails","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2606.28153","citing_title":"Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models","ref_index":80,"is_internal_anchor":false},{"citing_arxiv_id":"2606.28153","citing_title":"Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2606.29887","citing_title":"SafePyramid: A Hierarchical Benchmark for In-context Policy Guardrailing","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2606.25034","citing_title":"Yuvion VL: A Multimodal Foundation Model for Adversarial Content and AI Safety","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2605.29659","citing_title":"Opir: Efficient Multi-Task Safety Classification for Toxicity, Jailbreaks, Hate Speech, and Harmful Content","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2606.27632","citing_title":"Yuvion LLM: An Adversarially-Aware Large Language Model for Content And AI Safety","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2605.31073","citing_title":"ConsisGuard: Aligning Safety Deliberation with Policy Enforcement in LLM Guardrails","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00611","citing_title":"TRACE: Trajectory Risk-Aware Compression for Long-Horizon Agent Safety","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2502.01241","citing_title":"Peering Behind the Shield: Guardrail Identification in Large Language Models","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2605.15598","citing_title":"Compositional Jailbreaking: An Empirical Analysis of Mutator Chain Interactions in Aligned LLMs","ref_index":36,"is_internal_anchor":false},{"citing_arxiv_id":"2407.21772","citing_title":"ShieldGemma: Generative AI Content Moderation Based on Gemma","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2506.00166","citing_title":"Disentangled Safety Adapters Enable Efficient Guardrails and Flexible Inference-Time Alignment","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2406.18495","citing_title":"WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2604.25716","citing_title":"Cross-Lingual Jailbreak Detection via Semantic Codebooks","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2604.07655","citing_title":"Guardian-as-an-Advisor: Advancing Next-Generation Guardian Models for Trustworthy LLMs","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08608","citing_title":"Semantic Intent Fragmentation: A Single-Shot Compositional Attack on Multi-Agent AI Pipelines","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2501.03575","citing_title":"Cosmos World Foundation Model Platform for Physical AI","ref_index":56,"is_internal_anchor":false},{"citing_arxiv_id":"2604.18519","citing_title":"LLM Safety From Within: Detecting Harmful Content with Internal Representations","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.04572","citing_title":"From Parameter Dynamics to Risk Scoring : Quantifying Sample-Level Safety Degradation in LLM Fine-tuning","ref_index":4,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/YEJU657KWRV3SKWAOLTA2XSEFL","json":"https://pith.science/pith/YEJU657KWRV3SKWAOLTA2XSEFL.json","graph_json":"https://pith.science/api/pith-number/YEJU657KWRV3SKWAOLTA2XSEFL/graph.json","events_json":"https://pith.science/api/pith-number/YEJU657KWRV3SKWAOLTA2XSEFL/events.json","paper":"https://pith.science/paper/YEJU657K"},"agent_actions":{"view_html":"https://pith.science/pith/YEJU657KWRV3SKWAOLTA2XSEFL","download_json":"https://pith.science/pith/YEJU657KWRV3SKWAOLTA2XSEFL.json","view_paper":"https://pith.science/paper/YEJU657K","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2404.05993&json=true","fetch_graph":"https://pith.science/api/pith-number/YEJU657KWRV3SKWAOLTA2XSEFL/graph.json","fetch_events":"https://pith.science/api/pith-number/YEJU657KWRV3SKWAOLTA2XSEFL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/YEJU657KWRV3SKWAOLTA2XSEFL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/YEJU657KWRV3SKWAOLTA2XSEFL/action/storage_attestation","attest_author":"https://pith.science/pith/YEJU657KWRV3SKWAOLTA2XSEFL/action/author_attestation","sign_citation":"https://pith.science/pith/YEJU657KWRV3SKWAOLTA2XSEFL/action/citation_signature","submit_replication":"https://pith.science/pith/YEJU657KWRV3SKWAOLTA2XSEFL/action/replication_record"}},"created_at":"2026-07-05T09:05:54.291870+00:00","updated_at":"2026-07-05T09:05:54.291870+00:00"}