{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:6KPSY3BGXGESLNR4Z3PVMGIG4O","short_pith_number":"pith:6KPSY3BG","schema_version":"1.0","canonical_sha256":"f29f2c6c26b98925b63ccedf561906e3bc09ed6d442d6352bd38a101525d05c9","source":{"kind":"arxiv","id":"2506.08837","version":3},"attestation_state":"computed","paper":{"title":"Design Patterns for Securing LLM Agents against Prompt Injections","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CR"],"primary_cat":"cs.LG","authors_text":"Ana-Maria Cre\\c{t}u, Andrew Paverd, Beat Buesser, Daniel Dobos, Daniel Fabian, Daniel Naeff, David Froelicher, Edoardo Debenedetti, Ezinwanne Ozoani, Florian Tram\\`er, Kathrin Grosse, Luca Beurer-Kellner, Marc Fischer, V\\'aclav Volhejn","submitted_at":"2025-06-10T14:23:55Z","abstract_excerpt":"As AI agents powered by Large Language Models (LLMs) become increasingly versatile and capable of addressing a broad spectrum of tasks, ensuring their security has become a critical challenge. Among the most pressing threats are prompt injection attacks, which exploit the agent's resilience on natural language inputs -- an especially dangerous threat when agents are granted tool access or handle sensitive information. In this work, we propose a set of principled design patterns for building AI agents with provable resistance to prompt injection. We systematically analyze these patterns, discus"},"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":"2506.08837","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-06-10T14:23:55Z","cross_cats_sorted":["cs.CR"],"title_canon_sha256":"82d9df10995c82de8f19015c9f4d23cc776229ef18a43dd748d37133968d774d","abstract_canon_sha256":"220eb6dd758bfcccbbfaa85809b81d7e33d52c3b157cd909a7758d836e69b8f2"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:28:17.217912Z","signature_b64":"lt6NRFhqdtM4kl+aIrnsqVndIa6yaFRrPktfVtA5ndJ2zI7myRGIQKJRER80bT+FxulJ+ZcrOayWYgSEvxarBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f29f2c6c26b98925b63ccedf561906e3bc09ed6d442d6352bd38a101525d05c9","last_reissued_at":"2026-07-05T11:28:17.217413Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:28:17.217413Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Design Patterns for Securing LLM Agents against Prompt Injections","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CR"],"primary_cat":"cs.LG","authors_text":"Ana-Maria Cre\\c{t}u, Andrew Paverd, Beat Buesser, Daniel Dobos, Daniel Fabian, Daniel Naeff, David Froelicher, Edoardo Debenedetti, Ezinwanne Ozoani, Florian Tram\\`er, Kathrin Grosse, Luca Beurer-Kellner, Marc Fischer, V\\'aclav Volhejn","submitted_at":"2025-06-10T14:23:55Z","abstract_excerpt":"As AI agents powered by Large Language Models (LLMs) become increasingly versatile and capable of addressing a broad spectrum of tasks, ensuring their security has become a critical challenge. Among the most pressing threats are prompt injection attacks, which exploit the agent's resilience on natural language inputs -- an especially dangerous threat when agents are granted tool access or handle sensitive information. In this work, we propose a set of principled design patterns for building AI agents with provable resistance to prompt injection. We systematically analyze these patterns, discus"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.08837","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/2506.08837/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":"2506.08837","created_at":"2026-07-05T11:28:17.217467+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.08837v3","created_at":"2026-07-05T11:28:17.217467+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.08837","created_at":"2026-07-05T11:28:17.217467+00:00"},{"alias_kind":"pith_short_12","alias_value":"6KPSY3BGXGES","created_at":"2026-07-05T11:28:17.217467+00:00"},{"alias_kind":"pith_short_16","alias_value":"6KPSY3BGXGESLNR4","created_at":"2026-07-05T11:28:17.217467+00:00"},{"alias_kind":"pith_short_8","alias_value":"6KPSY3BG","created_at":"2026-07-05T11:28:17.217467+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":21,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.08147","citing_title":"Prismata: Confining Cross-Site Prompt Injection in Web Agents","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2606.26627","citing_title":"Agents That Know Too Much: A Data-Centric Survey of Privacy in LLM Agents","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2606.26479","citing_title":"Adaptive Evaluation of Out-of-Band Defenses Against Prompt Injection in LLM Agents","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2606.23277","citing_title":"GIF: Locally Sound Geometric Information Flow Control for LLMs","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2606.22504","citing_title":"Lingering Authority: Revocable Resource-and-Effect Capabilities for Coding Agents","ref_index":5,"is_internal_anchor":false},{"citing_arxiv_id":"2607.01510","citing_title":"Janus: a Playground for User-Involved Agentic Permission Management","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10322","citing_title":"Game-Theoretic Multi-Agent Control for Robust Contextual Reasoning in LLMs","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05679","citing_title":"Data Flow Control: Data Safety Policies for AI Agents","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02080","citing_title":"Agentic-J: An AI Agent for Biological Microscopy Image Analysis","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02668","citing_title":"What You Approve Is What Executes: Consent Integrity for Black-Box LLM Agents","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10749","citing_title":"Toward Secure LLM Agents: Threat Surfaces, Attacks, Defenses, and Evaluation","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2512.06556","citing_title":"Semantic Attacks on Tool-Augmented LLMs: Securing the Model Context Protocol Against Descriptor-Level Manipulation","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18784","citing_title":"The Insurability Frontier of AI Risk: Mapping Threats to Affirmative Coverage, Silent Exposures, and Exclusions","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01970","citing_title":"Trojan Hippo: Weaponizing Agent Memory for Data Exfiltration","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2510.23883","citing_title":"Agentic AI Security: Threats, Defenses, Evaluation, and Open Challenges","ref_index":59,"is_internal_anchor":false},{"citing_arxiv_id":"2504.19793","citing_title":"Prompt Injection Attack to Tool Selection in LLM Agents","ref_index":73,"is_internal_anchor":false},{"citing_arxiv_id":"2603.00991","citing_title":"Tracking Capabilities for Safer Agents","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2605.14290","citing_title":"Web Agents Should Adopt the Plan-Then-Execute Paradigm","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11868","citing_title":"IPI-proxy: An Intercepting Proxy for Red-Teaming Web-Browsing AI Agents Against Indirect Prompt Injection","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00081","citing_title":"Alignment Contracts for Agentic Security Systems","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2604.15368","citing_title":"LogJack: Indirect Prompt Injection Through Cloud Logs Against LLM Debugging Agents","ref_index":1,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6KPSY3BGXGESLNR4Z3PVMGIG4O","json":"https://pith.science/pith/6KPSY3BGXGESLNR4Z3PVMGIG4O.json","graph_json":"https://pith.science/api/pith-number/6KPSY3BGXGESLNR4Z3PVMGIG4O/graph.json","events_json":"https://pith.science/api/pith-number/6KPSY3BGXGESLNR4Z3PVMGIG4O/events.json","paper":"https://pith.science/paper/6KPSY3BG"},"agent_actions":{"view_html":"https://pith.science/pith/6KPSY3BGXGESLNR4Z3PVMGIG4O","download_json":"https://pith.science/pith/6KPSY3BGXGESLNR4Z3PVMGIG4O.json","view_paper":"https://pith.science/paper/6KPSY3BG","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.08837&json=true","fetch_graph":"https://pith.science/api/pith-number/6KPSY3BGXGESLNR4Z3PVMGIG4O/graph.json","fetch_events":"https://pith.science/api/pith-number/6KPSY3BGXGESLNR4Z3PVMGIG4O/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6KPSY3BGXGESLNR4Z3PVMGIG4O/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6KPSY3BGXGESLNR4Z3PVMGIG4O/action/storage_attestation","attest_author":"https://pith.science/pith/6KPSY3BGXGESLNR4Z3PVMGIG4O/action/author_attestation","sign_citation":"https://pith.science/pith/6KPSY3BGXGESLNR4Z3PVMGIG4O/action/citation_signature","submit_replication":"https://pith.science/pith/6KPSY3BGXGESLNR4Z3PVMGIG4O/action/replication_record"}},"created_at":"2026-07-05T11:28:17.217467+00:00","updated_at":"2026-07-05T11:28:17.217467+00:00"}