{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:HRLXLYZ2PM4E4ZYP4SQ6VZYL4U","short_pith_number":"pith:HRLXLYZ2","schema_version":"1.0","canonical_sha256":"3c5775e33a7b384e670fe4a1eae70be508d9494b6d3fe7c16658b27889544671","source":{"kind":"arxiv","id":"2403.09539","version":3},"attestation_state":"computed","paper":{"title":"Logits of API-Protected LLMs Leak Proprietary Information","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CR","cs.LG"],"primary_cat":"cs.CL","authors_text":"Matthew Finlayson, Swabha Swayamdipta, Xiang Ren","submitted_at":"2024-03-14T16:27:49Z","abstract_excerpt":"Large language model (LLM) providers often hide the architectural details and parameters of their proprietary models by restricting public access to a limited API. In this work we show that, with only a conservative assumption about the model architecture, it is possible to learn a surprisingly large amount of non-public information about an API-protected LLM from a relatively small number of API queries (e.g., costing under $1000 USD for OpenAI's gpt-3.5-turbo). Our findings are centered on one key observation: most modern LLMs suffer from a softmax bottleneck, which restricts the model outpu"},"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":"2403.09539","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-03-14T16:27:49Z","cross_cats_sorted":["cs.AI","cs.CR","cs.LG"],"title_canon_sha256":"2f79c821b87f98a4a0e7f04e04091bfad7b06b460bc3b470dd30579562a54828","abstract_canon_sha256":"48ad2834e0d8573ca04130105f3776200f9433c41b306e5a8cbb5b0e11c93e63"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:32:42.955246Z","signature_b64":"Oh7GDZGBjRcRSJNtZ1Jz9E6VlmItWEi2+o+rRhJTS6wDHZCqT/LQjMX7fwJW7Q0Ej/D1htsIC32x9s1gEbPYCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3c5775e33a7b384e670fe4a1eae70be508d9494b6d3fe7c16658b27889544671","last_reissued_at":"2026-07-05T09:32:42.954249Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:32:42.954249Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Logits of API-Protected LLMs Leak Proprietary Information","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CR","cs.LG"],"primary_cat":"cs.CL","authors_text":"Matthew Finlayson, Swabha Swayamdipta, Xiang Ren","submitted_at":"2024-03-14T16:27:49Z","abstract_excerpt":"Large language model (LLM) providers often hide the architectural details and parameters of their proprietary models by restricting public access to a limited API. In this work we show that, with only a conservative assumption about the model architecture, it is possible to learn a surprisingly large amount of non-public information about an API-protected LLM from a relatively small number of API queries (e.g., costing under $1000 USD for OpenAI's gpt-3.5-turbo). Our findings are centered on one key observation: most modern LLMs suffer from a softmax bottleneck, which restricts the model outpu"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2403.09539","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/2403.09539/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":"2403.09539","created_at":"2026-07-05T09:32:42.954712+00:00"},{"alias_kind":"arxiv_version","alias_value":"2403.09539v3","created_at":"2026-07-05T09:32:42.954712+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2403.09539","created_at":"2026-07-05T09:32:42.954712+00:00"},{"alias_kind":"pith_short_12","alias_value":"HRLXLYZ2PM4E","created_at":"2026-07-05T09:32:42.954712+00:00"},{"alias_kind":"pith_short_16","alias_value":"HRLXLYZ2PM4E4ZYP","created_at":"2026-07-05T09:32:42.954712+00:00"},{"alias_kind":"pith_short_8","alias_value":"HRLXLYZ2","created_at":"2026-07-05T09:32:42.954712+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.22019","citing_title":"Channel Location Constrains the Auditability of Subliminal Learning","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06854","citing_title":"The Geometry of Last-Layer Model Stealing","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2406.11354","citing_title":"Preserving Knowledge in Large Language Model with Model-Agnostic Self-Decompression","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10407","citing_title":"Identified-Set Geometry of Distributional Model Extraction under Top-$K$ Censored API Access","ref_index":8,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U","json":"https://pith.science/pith/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U.json","graph_json":"https://pith.science/api/pith-number/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U/graph.json","events_json":"https://pith.science/api/pith-number/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U/events.json","paper":"https://pith.science/paper/HRLXLYZ2"},"agent_actions":{"view_html":"https://pith.science/pith/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U","download_json":"https://pith.science/pith/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U.json","view_paper":"https://pith.science/paper/HRLXLYZ2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2403.09539&json=true","fetch_graph":"https://pith.science/api/pith-number/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U/graph.json","fetch_events":"https://pith.science/api/pith-number/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U/action/timestamp_anchor","attest_storage":"https://pith.science/pith/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U/action/storage_attestation","attest_author":"https://pith.science/pith/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U/action/author_attestation","sign_citation":"https://pith.science/pith/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U/action/citation_signature","submit_replication":"https://pith.science/pith/HRLXLYZ2PM4E4ZYP4SQ6VZYL4U/action/replication_record"}},"created_at":"2026-07-05T09:32:42.954712+00:00","updated_at":"2026-07-05T09:32:42.954712+00:00"}