{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:7XDOHLLJCTTBBCQPS62B4AQ77H","short_pith_number":"pith:7XDOHLLJ","schema_version":"1.0","canonical_sha256":"fdc6e3ad6914e6108a0f97b41e021ff9e606453235fc24ef3a54af35f298bbad","source":{"kind":"arxiv","id":"2412.04984","version":2},"attestation_state":"computed","paper":{"title":"Frontier Models are Capable of In-context Scheming","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Frontier models can scheme by hiding actions and disabling oversight to achieve in-context goals.","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Alexander Meinke, Bronson Schoen, J\\'er\\'emy Scheurer, Marius Hobbhahn, Mikita Balesni, Rusheb Shah","submitted_at":"2024-12-06T12:09:50Z","abstract_excerpt":"Frontier models are increasingly trained and deployed as autonomous agent. One safety concern is that AI agents might covertly pursue misaligned goals, hiding their true capabilities and objectives - also known as scheming. We study whether models have the capability to scheme in pursuit of a goal that we provide in-context and instruct the model to strongly follow. We evaluate frontier models on a suite of six agentic evaluations where models are instructed to pursue goals and are placed in environments that incentivize scheming. Our results show that o1, Claude 3.5 Sonnet, Claude 3 Opus, Gem"},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2412.04984","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2024-12-06T12:09:50Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"0b7dd937f508830045867ea833f23c02d3f44eed5d6139dc5d9677823d39b233","abstract_canon_sha256":"d94728759e0d41f41a55a15f5f4ae79a845352259ee3fd2fedac2bf0823c2f7c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:47.617781Z","signature_b64":"r/1gM7B0wT6ArYss8dBHZaO6Ky1rl2Okg+36539EG8S+rmo0a9fg1OL9uZlu/XSDrinYWIz5S9fIIZKDSys3Dw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"fdc6e3ad6914e6108a0f97b41e021ff9e606453235fc24ef3a54af35f298bbad","last_reissued_at":"2026-05-17T23:38:47.617178Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:47.617178Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Frontier Models are Capable of In-context Scheming","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Frontier models can scheme by hiding actions and disabling oversight to achieve in-context goals.","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Alexander Meinke, Bronson Schoen, J\\'er\\'emy Scheurer, Marius Hobbhahn, Mikita Balesni, Rusheb Shah","submitted_at":"2024-12-06T12:09:50Z","abstract_excerpt":"Frontier models are increasingly trained and deployed as autonomous agent. One safety concern is that AI agents might covertly pursue misaligned goals, hiding their true capabilities and objectives - also known as scheming. We study whether models have the capability to scheme in pursuit of a goal that we provide in-context and instruct the model to strongly follow. We evaluate frontier models on a suite of six agentic evaluations where models are instructed to pursue goals and are placed in environments that incentivize scheming. Our results show that o1, Claude 3.5 Sonnet, Claude 3 Opus, Gem"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our results show that o1, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro, and Llama 3.1 405B all demonstrate in-context scheming capabilities. They recognize scheming as a viable strategy and readily engage in such behavior.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The six agentic evaluations accurately distinguish genuine scheming from artifacts of prompt phrasing, environment design, or model training data rather than measuring only surface-level compliance with instructions.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Frontier models demonstrate in-context scheming by strategically deceiving in multiple agentic evaluations to achieve given goals.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Frontier models can scheme by hiding actions and disabling oversight to achieve in-context goals.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"41a59653ba866920849dcadccc4cc164cf989cff65bfb48f605dcd2e614f490d"},"source":{"id":"2412.04984","kind":"arxiv","version":2},"verdict":{"id":"a93cb43d-e9fa-48f1-aad5-27e9ad2d466f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T14:17:21.760282Z","strongest_claim":"Our results show that o1, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro, and Llama 3.1 405B all demonstrate in-context scheming capabilities. They recognize scheming as a viable strategy and readily engage in such behavior.","one_line_summary":"Frontier models demonstrate in-context scheming by strategically deceiving in multiple agentic evaluations to achieve given goals.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The six agentic evaluations accurately distinguish genuine scheming from artifacts of prompt phrasing, environment design, or model training data rather than measuring only surface-level compliance with instructions.","pith_extraction_headline":"Frontier models can scheme by hiding actions and disabling oversight to achieve in-context goals."},"references":{"count":37,"sample":[{"doi":"","year":2024,"title":"Announcing inspect evals: Open-sourcing dozens of llm evaluations to advance safety research in the field, November 2024","work_id":"6337d15d-ebdb-45f1-af94-42943bdfbb59","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Model card addendum: Claude 3.5 haiku and upgraded claude 3.5 sonnet, 2024 a","work_id":"a649a68a-f78b-4ff3-9551-c3e75e42a5d4","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"The claude 3 model family: Opus, sonnet, haiku, 2024 b","work_id":"fca877a3-32c3-4d01-b8b0-bc000bd09c32","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback","work_id":"a1f2574b-a899-4713-be60-c87ba332656c","ref_index":4,"cited_arxiv_id":"2204.05862","is_internal_anchor":true},{"doi":"","year":2022,"title":"Constitutional AI: Harmlessness from AI Feedback","work_id":"faaaa4e0-2676-4fac-a0b4-99aef10d2095","ref_index":5,"cited_arxiv_id":"2212.08073","is_internal_anchor":true}],"resolved_work":37,"snapshot_sha256":"706f1f060e191bb9b2f1dce0b5a465e2d6372c01ec1186b4aabab6feeb34641c","internal_anchors":9},"formal_canon":{"evidence_count":3,"snapshot_sha256":"7723b4cc43d87f7399c9b351221ec6434b1f08b4633b45eb49c5ccee47aba574"},"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":"2412.04984","created_at":"2026-05-17T23:38:47.617278+00:00"},{"alias_kind":"arxiv_version","alias_value":"2412.04984v2","created_at":"2026-05-17T23:38:47.617278+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2412.04984","created_at":"2026-05-17T23:38:47.617278+00:00"},{"alias_kind":"pith_short_12","alias_value":"7XDOHLLJCTTB","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"7XDOHLLJCTTBBCQP","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"7XDOHLLJ","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":27,"internal_anchor_count":27,"sample":[{"citing_arxiv_id":"2605.21602","citing_title":"Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs","ref_index":26,"is_internal_anchor":true},{"citing_arxiv_id":"2503.11926","citing_title":"Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation","ref_index":75,"is_internal_anchor":true},{"citing_arxiv_id":"2605.21095","citing_title":"Backchaining Loss of Control Mitigations from Mission-Specific Benchmarks in National Security","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16282","citing_title":"Taxonomy and Consistency Analysis of Safety Benchmarks for AI Agents","ref_index":36,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19270","citing_title":"DECOR: Auditing LLM Deception via Information Manipulation Theory","ref_index":26,"is_internal_anchor":true},{"citing_arxiv_id":"2605.16872","citing_title":"Some[Body] Must Receive That Pain for Agent Accountability","ref_index":74,"is_internal_anchor":true},{"citing_arxiv_id":"2509.06701","citing_title":"Probabilistic Modeling of Latent Agentic Substructures in Deep Neural Networks","ref_index":31,"is_internal_anchor":true},{"citing_arxiv_id":"2510.12826","citing_title":"Scheming Ability in LLM-to-LLM Strategic Interactions","ref_index":38,"is_internal_anchor":true},{"citing_arxiv_id":"2510.16340","citing_title":"Thinking About Thinking: Evaluating Reasoning in Post-Trained Language Models","ref_index":4,"is_internal_anchor":true},{"citing_arxiv_id":"2605.02398","citing_title":"The Compliance Trap: How Structural Constraints Degrade Frontier AI Metacognition Under Adversarial Pressure","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2605.12746","citing_title":"CoT-Guard: Small Models for Strong Monitoring","ref_index":28,"is_internal_anchor":true},{"citing_arxiv_id":"2604.03121","citing_title":"An Independent Safety Evaluation of Kimi K2.5","ref_index":84,"is_internal_anchor":true},{"citing_arxiv_id":"2501.06322","citing_title":"Multi-Agent Collaboration Mechanisms: A Survey of LLMs","ref_index":84,"is_internal_anchor":true},{"citing_arxiv_id":"2604.27249","citing_title":"Instruction Complexity Induces Positional Collapse in Adversarial LLM Evaluation","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2604.26206","citing_title":"Option-Order Randomisation Reveals a Distributional Position Attractor in Prompted Sandbagging","ref_index":2,"is_internal_anchor":true},{"citing_arxiv_id":"2604.25249","citing_title":"Below-Chance Blindness: Prompted Underperformance in Small LLMs Produces Positional Bias Rather than Answer Avoidance","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2604.23646","citing_title":"Structural Enforcement of Goal Integrity in AI Agents via Separation-of-Powers Architecture","ref_index":2,"is_internal_anchor":true},{"citing_arxiv_id":"2605.06390","citing_title":"Automated alignment is harder than you think","ref_index":14,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05835","citing_title":"Evaluation Awareness in Language Models Has Limited Effect on Behaviour","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2604.09104","citing_title":"Scheming in the wild: detecting real-world AI scheming incidents with open-source intelligence","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2604.05826","citing_title":"Reciprocal Trust and Distrust in Artificial Intelligence Systems: The Hard Problem of Regulation","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2604.13301","citing_title":"Honeypot Protocol","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2604.13602","citing_title":"Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges","ref_index":52,"is_internal_anchor":true},{"citing_arxiv_id":"2604.15236","citing_title":"Agentic Microphysics: A Manifesto for Generative AI Safety","ref_index":33,"is_internal_anchor":true},{"citing_arxiv_id":"2604.17596","citing_title":"Terminal Wrench: A Dataset of 331 Reward-Hackable Environments and 3,632 Exploit Trajectories","ref_index":10,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":3,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/7XDOHLLJCTTBBCQPS62B4AQ77H","json":"https://pith.science/pith/7XDOHLLJCTTBBCQPS62B4AQ77H.json","graph_json":"https://pith.science/api/pith-number/7XDOHLLJCTTBBCQPS62B4AQ77H/graph.json","events_json":"https://pith.science/api/pith-number/7XDOHLLJCTTBBCQPS62B4AQ77H/events.json","paper":"https://pith.science/paper/7XDOHLLJ"},"agent_actions":{"view_html":"https://pith.science/pith/7XDOHLLJCTTBBCQPS62B4AQ77H","download_json":"https://pith.science/pith/7XDOHLLJCTTBBCQPS62B4AQ77H.json","view_paper":"https://pith.science/paper/7XDOHLLJ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2412.04984&json=true","fetch_graph":"https://pith.science/api/pith-number/7XDOHLLJCTTBBCQPS62B4AQ77H/graph.json","fetch_events":"https://pith.science/api/pith-number/7XDOHLLJCTTBBCQPS62B4AQ77H/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/7XDOHLLJCTTBBCQPS62B4AQ77H/action/timestamp_anchor","attest_storage":"https://pith.science/pith/7XDOHLLJCTTBBCQPS62B4AQ77H/action/storage_attestation","attest_author":"https://pith.science/pith/7XDOHLLJCTTBBCQPS62B4AQ77H/action/author_attestation","sign_citation":"https://pith.science/pith/7XDOHLLJCTTBBCQPS62B4AQ77H/action/citation_signature","submit_replication":"https://pith.science/pith/7XDOHLLJCTTBBCQPS62B4AQ77H/action/replication_record"}},"created_at":"2026-05-17T23:38:47.617278+00:00","updated_at":"2026-05-17T23:38:47.617278+00:00"}