{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:EQDVWM4N542VFVJNOCOEFZ6WG5","short_pith_number":"pith:EQDVWM4N","schema_version":"1.0","canonical_sha256":"24075b338def3552d52d709c42e7d637429cc3ab4d1ccedfd6a6974c3869e6d1","source":{"kind":"arxiv","id":"2505.03439","version":1},"attestation_state":"computed","paper":{"title":"The Steganographic Potentials of Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CR","cs.LG"],"primary_cat":"cs.AI","authors_text":"Artem Karpov, Natalia Perez-Campanero, Seong Hah Cho, Tinuade Adeleke","submitted_at":"2025-05-06T11:25:52Z","abstract_excerpt":"The potential for large language models (LLMs) to hide messages within plain text (steganography) poses a challenge to detection and thwarting of unaligned AI agents, and undermines faithfulness of LLMs reasoning. We explore the steganographic capabilities of LLMs fine-tuned via reinforcement learning (RL) to: (1) develop covert encoding schemes, (2) engage in steganography when prompted, and (3) utilize steganography in realistic scenarios where hidden reasoning is likely, but not prompted. In these scenarios, we detect the intention of LLMs to hide their reasoning as well as their steganogra"},"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":"2505.03439","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-05-06T11:25:52Z","cross_cats_sorted":["cs.CR","cs.LG"],"title_canon_sha256":"a5bfe55fbdd0669fb63f9e90873e28991d83b5db3a393a225144fa0fa3db83b8","abstract_canon_sha256":"639db40c7f98ca36aba2f2d25a697971a66270af0f80039189d2a06010da581c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:59:16.284676Z","signature_b64":"59186jriHltZXzEm7eC/H65ZeaAuNMQuEhTNmZJS1gNKa9X0zv35XCunRA6JA+1Llqf3YP1O5Xwp1K2POFFqCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"24075b338def3552d52d709c42e7d637429cc3ab4d1ccedfd6a6974c3869e6d1","last_reissued_at":"2026-07-05T10:59:16.284243Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:59:16.284243Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Steganographic Potentials of Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CR","cs.LG"],"primary_cat":"cs.AI","authors_text":"Artem Karpov, Natalia Perez-Campanero, Seong Hah Cho, Tinuade Adeleke","submitted_at":"2025-05-06T11:25:52Z","abstract_excerpt":"The potential for large language models (LLMs) to hide messages within plain text (steganography) poses a challenge to detection and thwarting of unaligned AI agents, and undermines faithfulness of LLMs reasoning. We explore the steganographic capabilities of LLMs fine-tuned via reinforcement learning (RL) to: (1) develop covert encoding schemes, (2) engage in steganography when prompted, and (3) utilize steganography in realistic scenarios where hidden reasoning is likely, but not prompted. In these scenarios, we detect the intention of LLMs to hide their reasoning as well as their steganogra"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.03439","kind":"arxiv","version":1},"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/2505.03439/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":"2505.03439","created_at":"2026-07-05T10:59:16.284299+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.03439v1","created_at":"2026-07-05T10:59:16.284299+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.03439","created_at":"2026-07-05T10:59:16.284299+00:00"},{"alias_kind":"pith_short_12","alias_value":"EQDVWM4N542V","created_at":"2026-07-05T10:59:16.284299+00:00"},{"alias_kind":"pith_short_16","alias_value":"EQDVWM4N542VFVJN","created_at":"2026-07-05T10:59:16.284299+00:00"},{"alias_kind":"pith_short_8","alias_value":"EQDVWM4N","created_at":"2026-07-05T10:59:16.284299+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.28166","citing_title":"Tandem Reinforcement Learning with Verifiable Rewards","ref_index":8,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/EQDVWM4N542VFVJNOCOEFZ6WG5","json":"https://pith.science/pith/EQDVWM4N542VFVJNOCOEFZ6WG5.json","graph_json":"https://pith.science/api/pith-number/EQDVWM4N542VFVJNOCOEFZ6WG5/graph.json","events_json":"https://pith.science/api/pith-number/EQDVWM4N542VFVJNOCOEFZ6WG5/events.json","paper":"https://pith.science/paper/EQDVWM4N"},"agent_actions":{"view_html":"https://pith.science/pith/EQDVWM4N542VFVJNOCOEFZ6WG5","download_json":"https://pith.science/pith/EQDVWM4N542VFVJNOCOEFZ6WG5.json","view_paper":"https://pith.science/paper/EQDVWM4N","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.03439&json=true","fetch_graph":"https://pith.science/api/pith-number/EQDVWM4N542VFVJNOCOEFZ6WG5/graph.json","fetch_events":"https://pith.science/api/pith-number/EQDVWM4N542VFVJNOCOEFZ6WG5/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/EQDVWM4N542VFVJNOCOEFZ6WG5/action/timestamp_anchor","attest_storage":"https://pith.science/pith/EQDVWM4N542VFVJNOCOEFZ6WG5/action/storage_attestation","attest_author":"https://pith.science/pith/EQDVWM4N542VFVJNOCOEFZ6WG5/action/author_attestation","sign_citation":"https://pith.science/pith/EQDVWM4N542VFVJNOCOEFZ6WG5/action/citation_signature","submit_replication":"https://pith.science/pith/EQDVWM4N542VFVJNOCOEFZ6WG5/action/replication_record"}},"created_at":"2026-07-05T10:59:16.284299+00:00","updated_at":"2026-07-05T10:59:16.284299+00:00"}