{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:63EZNDZX5H4HA3UCH6ZYPJDYDI","short_pith_number":"pith:63EZNDZX","schema_version":"1.0","canonical_sha256":"f6c9968f37e9f8706e823fb387a4781a3076e229c792b40b8b6be573b913d898","source":{"kind":"arxiv","id":"2403.06634","version":2},"attestation_state":"computed","paper":{"title":"Stealing Part of a Production Language Model","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CR","authors_text":"A. Feder Cooper, Arthur Conmy, Daniel Paleka, David Rolnick, Eric Wallace, Florian Tram\\`er, Itay Yona, Jonathan Hayase, Katherine Lee, Krishnamurthy Dj Dvijotham, Matthew Jagielski, Milad Nasr, Nicholas Carlini, Thomas Steinke","submitted_at":"2024-03-11T11:46:12Z","abstract_excerpt":"We introduce the first model-stealing attack that extracts precise, nontrivial information from black-box production language models like OpenAI's ChatGPT or Google's PaLM-2. Specifically, our attack recovers the embedding projection layer (up to symmetries) of a transformer model, given typical API access. For under \\$20 USD, our attack extracts the entire projection matrix of OpenAI's Ada and Babbage language models. We thereby confirm, for the first time, that these black-box models have a hidden dimension of 1024 and 2048, respectively. We also recover the exact hidden dimension size of th"},"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":"2403.06634","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CR","submitted_at":"2024-03-11T11:46:12Z","cross_cats_sorted":[],"title_canon_sha256":"db55d9d33349fbf3b21083fb0a067314669859b1d3c2c849dc318857df2ed095","abstract_canon_sha256":"e5c553ad19dd7448f3b10292bb0080072f337f8532df4bda75072523266e6e03"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:41:43.872061Z","signature_b64":"ZEG5CAO2bykkSHulYpXWBAnTS+KnpBzC+racWWhRDJjEAm+NIVdtuadFGcXNFIe3cnRu9W047dcMAgLA5aENBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f6c9968f37e9f8706e823fb387a4781a3076e229c792b40b8b6be573b913d898","last_reissued_at":"2026-07-05T08:41:43.871550Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:41:43.871550Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Stealing Part of a Production Language Model","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CR","authors_text":"A. Feder Cooper, Arthur Conmy, Daniel Paleka, David Rolnick, Eric Wallace, Florian Tram\\`er, Itay Yona, Jonathan Hayase, Katherine Lee, Krishnamurthy Dj Dvijotham, Matthew Jagielski, Milad Nasr, Nicholas Carlini, Thomas Steinke","submitted_at":"2024-03-11T11:46:12Z","abstract_excerpt":"We introduce the first model-stealing attack that extracts precise, nontrivial information from black-box production language models like OpenAI's ChatGPT or Google's PaLM-2. Specifically, our attack recovers the embedding projection layer (up to symmetries) of a transformer model, given typical API access. For under \\$20 USD, our attack extracts the entire projection matrix of OpenAI's Ada and Babbage language models. We thereby confirm, for the first time, that these black-box models have a hidden dimension of 1024 and 2048, respectively. We also recover the exact hidden dimension size of th"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2403.06634","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/2403.06634/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":"2403.06634","created_at":"2026-07-05T08:41:43.871611+00:00"},{"alias_kind":"arxiv_version","alias_value":"2403.06634v2","created_at":"2026-07-05T08:41:43.871611+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2403.06634","created_at":"2026-07-05T08:41:43.871611+00:00"},{"alias_kind":"pith_short_12","alias_value":"63EZNDZX5H4H","created_at":"2026-07-05T08:41:43.871611+00:00"},{"alias_kind":"pith_short_16","alias_value":"63EZNDZX5H4HA3UC","created_at":"2026-07-05T08:41:43.871611+00:00"},{"alias_kind":"pith_short_8","alias_value":"63EZNDZX","created_at":"2026-07-05T08:41:43.871611+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":12,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.22019","citing_title":"Channel Location Constrains the Auditability of Subliminal Learning","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2607.01313","citing_title":"Black-Box Inference of LLM Architectural Properties with Restrictive API Access","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2606.17358","citing_title":"OTRO: Oblivious Tokenization Path with Square-Root ORAM","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10091","citing_title":"SoK: Colluding Adversaries in Machine Learning Pipelines","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06854","citing_title":"The Geometry of Last-Layer Model Stealing","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2606.32008","citing_title":"Surrogate Fidelity: When Can Open LLMs Explain Closed Ones?","ref_index":56,"is_internal_anchor":false},{"citing_arxiv_id":"2605.30454","citing_title":"The Surface You Test Is Not the Surface That Breaks","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2509.25448","citing_title":"Fingerprinting LLMs via Prompt Injection","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2507.02850","citing_title":"LLM Hypnosis: Exploiting User Feedback for Unauthorized Knowledge Injection to All Users","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11170","citing_title":"Unlearning with Asymmetric Sources: Improved Unlearning-Utility Trade-off with Public Data","ref_index":85,"is_internal_anchor":false},{"citing_arxiv_id":"2604.23338","citing_title":"A Systematic Survey of Security Threats and Defenses in LLM-Based AI Agents: A Layered Attack Surface Framework","ref_index":65,"is_internal_anchor":false},{"citing_arxiv_id":"2605.04901","citing_title":"On the (In-)Security of the Shuffling Defense in the Transformer Secure Inference","ref_index":163,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/63EZNDZX5H4HA3UCH6ZYPJDYDI","json":"https://pith.science/pith/63EZNDZX5H4HA3UCH6ZYPJDYDI.json","graph_json":"https://pith.science/api/pith-number/63EZNDZX5H4HA3UCH6ZYPJDYDI/graph.json","events_json":"https://pith.science/api/pith-number/63EZNDZX5H4HA3UCH6ZYPJDYDI/events.json","paper":"https://pith.science/paper/63EZNDZX"},"agent_actions":{"view_html":"https://pith.science/pith/63EZNDZX5H4HA3UCH6ZYPJDYDI","download_json":"https://pith.science/pith/63EZNDZX5H4HA3UCH6ZYPJDYDI.json","view_paper":"https://pith.science/paper/63EZNDZX","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2403.06634&json=true","fetch_graph":"https://pith.science/api/pith-number/63EZNDZX5H4HA3UCH6ZYPJDYDI/graph.json","fetch_events":"https://pith.science/api/pith-number/63EZNDZX5H4HA3UCH6ZYPJDYDI/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/63EZNDZX5H4HA3UCH6ZYPJDYDI/action/timestamp_anchor","attest_storage":"https://pith.science/pith/63EZNDZX5H4HA3UCH6ZYPJDYDI/action/storage_attestation","attest_author":"https://pith.science/pith/63EZNDZX5H4HA3UCH6ZYPJDYDI/action/author_attestation","sign_citation":"https://pith.science/pith/63EZNDZX5H4HA3UCH6ZYPJDYDI/action/citation_signature","submit_replication":"https://pith.science/pith/63EZNDZX5H4HA3UCH6ZYPJDYDI/action/replication_record"}},"created_at":"2026-07-05T08:41:43.871611+00:00","updated_at":"2026-07-05T08:41:43.871611+00:00"}