{"paper":{"title":"Tracking Capabilities for Safer Agents","license":"http://creativecommons.org/licenses/by/4.0/","headline":"AI agents can generate capability-safe code with no significant loss in task performance while the type system blocks leaks and side effects.","cross_cats":["cs.PL"],"primary_cat":"cs.AI","authors_text":"Cao Nguyen Pham, Martin Odersky, Oliver Bra\\v{c}evac, Yaoyu Zhao, Yichen Xu","submitted_at":"2026-03-01T08:39:37Z","abstract_excerpt":"AI agents that interact with the real world through tool calls pose fundamental safety challenges: agents might leak private information, cause unintended side effects, or be manipulated through prompt injection. To address these challenges, we propose to put the agent in a programming-language-based \"safety harness\": instead of calling tools directly, agents express their intentions as code in a capability-safe language: Scala 3 with capture checking. Capabilities are program variables that regulate access to effects and resources of interest. Scala's type system tracks capabilities staticall"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our experiments show that agents can generate capability-safe code with no significant loss in task performance, while the type system reliably prevents unsafe behaviors such as information leakage and malicious side effects.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That large language models can reliably produce correct, well-typed capability-safe Scala code for arbitrary tasks and that the capture-checking system covers all relevant safety properties for real-world tool use.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"AI agents can generate code in a capability-safe Scala dialect that statically prevents information leakage and malicious side effects while preserving task performance.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"AI agents can generate capability-safe code with no significant loss in task performance while the type system blocks leaks and side effects.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"6e208500aed47fdce7d628df7fe59c8a5cb56a469e82e0503052603ccf98c057"},"source":{"id":"2603.00991","kind":"arxiv","version":2},"verdict":{"id":"3e56d15c-2814-4102-9310-ff6a21cc77b6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T18:28:22.252763Z","strongest_claim":"Our experiments show that agents can generate capability-safe code with no significant loss in task performance, while the type system reliably prevents unsafe behaviors such as information leakage and malicious side effects.","one_line_summary":"AI agents can generate code in a capability-safe Scala dialect that statically prevents information leakage and malicious side effects while preserving task performance.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That large language models can reliably produce correct, well-typed capability-safe Scala code for arbitrary tasks and that the capture-checking system covers all relevant safety properties for real-world tool use.","pith_extraction_headline":"AI agents can generate capability-safe code with no significant loss in task performance while the type system blocks leaks and side effects."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2603.00991/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":86,"sample":[{"doi":"","year":2025,"title":"Amazon Web Services. 2025. Bedrock AgentCore policy. Accessed: 2025-06-01. https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/policy.htm l (cit. on p. 8)","work_id":"b71d8ceb-53b3-4192-915f-565dddd6d58d","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Amazon Web Services. 2024. Cedar policy language. Accessed: 2025-06-01. https://www.cedarpolicy.com/ (cit. on p. 8)","work_id":"7abc0276-b639-4d05-b542-82c4e8b298b8","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Nada Amin, Samuel Grütter, Martin Odersky, Tiark Rompf, and Sandro Stucki","work_id":"629297b0-1156-4df8-abe8-4f28d9e087e2","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1007/978-3-319-30936-1_14","year":null,"title":"InA List of Successes That Can Change the World(Lecture Notes in Computer Science)","work_id":"5c49c691-648b-4cec-a23b-0d93d2e3180f","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2016,"title":"Concrete Problems in AI Safety","work_id":"c8d14fbe-6eab-464a-95b3-778aabd82fa3","ref_index":5,"cited_arxiv_id":"1606.06565","is_internal_anchor":true}],"resolved_work":86,"snapshot_sha256":"c6dce866aa7b6e697ae532916fd7847b1e399d79e0d098d90bdbd1c69e6e8064","internal_anchors":15},"formal_canon":{"evidence_count":2,"snapshot_sha256":"0568ba6fd24d1739ebce857bc66776f56a1380f103fd2964da428dbe3c240c8a"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}