{"paper":{"title":"Trojan Hippo: Weaponizing Agent Memory for Data Exfiltration","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A single untrusted tool call can plant a dormant payload in an agent's memory that later activates to exfiltrate sensitive user data.","cross_cats":["cs.AI"],"primary_cat":"cs.CR","authors_text":"Darya Kaviani, David Wagner, Debeshee Das, Florian Tram\\`er, Julien Piet, Luca Beurer-Kellner","submitted_at":"2026-05-03T17:07:20Z","abstract_excerpt":"Memory systems enable otherwise-stateless LLM agents to persist user information across sessions, but also introduce a new attack surface. We characterize the Trojan Hippo attack, a class of persistent memory attacks that operates in a more realistic threat model than prior memory poisoning work: the attacker plants a dormant payload into an agent's long-term memory via a single untrusted tool call (e.g., a crafted email), which activates only when the user later discusses sensitive topics such as finance, health, or identity, and exfiltrates high-value personal data to the attacker.\n  While a"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Trojan Hippo achieves up to 85-100% ASR against current frontier models from OpenAI and Google, with planted memories successfully activating even after 100 benign sessions.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The evaluation assumes that the four memory backends (explicit tool memory, agentic memory, RAG, and sliding-window context) and the OpenEvolve-based adaptive red-teaming accurately represent real-world deployed agent systems and that the threat model of a single untrusted tool call is realistic for attackers.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"The paper defines and evaluates Trojan Hippo attacks on LLM agent memory, showing 85-100% success in data exfiltration across backends and reduced rates with defenses at varying utility costs.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A single untrusted tool call can plant a dormant payload in an agent's memory that later activates to exfiltrate sensitive user data.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c107e90e2b1e07816a83e1f68d4695f039ce806e17d3bc0a165280654c5ef867"},"source":{"id":"2605.01970","kind":"arxiv","version":3},"verdict":{"id":"865f2c63-71be-4adf-bc4b-953e4f912623","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T17:26:07.958042Z","strongest_claim":"Trojan Hippo achieves up to 85-100% ASR against current frontier models from OpenAI and Google, with planted memories successfully activating even after 100 benign sessions.","one_line_summary":"The paper defines and evaluates Trojan Hippo attacks on LLM agent memory, showing 85-100% success in data exfiltration across backends and reduced rates with defenses at varying utility costs.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The evaluation assumes that the four memory backends (explicit tool memory, agentic memory, RAG, and sliding-window context) and the OpenEvolve-based adaptive red-teaming accurately represent real-world deployed agent systems and that the threat model of a single untrusted tool call is realistic for attackers.","pith_extraction_headline":"A single untrusted tool call can plant a dormant payload in an agent's memory that later activates to exfiltrate sensitive user data."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.01970/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T16:50:08.716171Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"72d287e6f2159dbc77248144fee92a0d4ca22a2992adcb442dc6813fcde158e1"},"references":{"count":104,"sample":[{"doi":"","year":2026,"title":"https://openai.com/index/scaling-ai-for- everyone/","work_id":"dc16c4c9-f5b7-4439-96da-a9e3cd7460a6","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Trustworthy agentic ai systems: a cross-layer review of architectures, threat models, and governance strategies for real-world deployment.F1000Research, 14(905):905, 2025","work_id":"2c481df5-6b5f-4b61-ac36-e4a1d030f6e0","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Burtsev, and Evgeny Burnaev","work_id":"5e87b079-06e6-4268-83b2-f336d186debd","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Claude memory","work_id":"62c00962-89b8-40c0-9d6f-af29f0d870a2","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Import your ChatGPT history to Claude","work_id":"d55175f9-7904-4bba-a413-45dd5bb6c634","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":104,"snapshot_sha256":"3dfe87db60f7cc033ca973637399230ccf7477a426fe125369ced4a2f6a04884","internal_anchors":19},"formal_canon":{"evidence_count":2,"snapshot_sha256":"3932c6654419eddb7dee52d0fa757737dd5f7f47e26d8aa900849bb13b8916f8"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}