{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:CPWBRHATZL3TN6PA6BA2UNH2FD","short_pith_number":"pith:CPWBRHAT","schema_version":"1.0","canonical_sha256":"13ec189c13caf736f9e0f041aa34fa28e16fe313a39df369e6f1e165f678f80e","source":{"kind":"arxiv","id":"2605.09033","version":3},"attestation_state":"computed","paper":{"title":"ShadowMerge: A Novel Poisoning Attack on Graph-Based Agent Memory via Relation-Channel Conflicts","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"ShadowMerge poisons graph-based agent memory by injecting relations that share the same query-activated anchor and channel as legitimate evidence.","cross_cats":["cs.AI"],"primary_cat":"cs.CR","authors_text":"Lingyun Peng, Shuyu Li, Tiantian Ji, Xinran Liu, Yang Luo, Yong Liu, Zifeng Kang","submitted_at":"2026-05-09T16:16:41Z","abstract_excerpt":"Graph-based agent memory is increasingly used in LLM agents to support structured long-term recall and multi-hop reasoning, but it also creates a new poisoning surface: an attacker can inject a crafted relation into graph memory so that it is later retrieved and influences agent behavior. Existing agent-memory poisoning attacks mainly target flat textual records and are ineffective in graph-based memory because malicious relations often fail to be extracted, merged into the target anchor neighborhood, or retrieved for the victim query.\n  We present SHADOWMERGE, a poisoning attack against graph"},"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":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.09033","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CR","submitted_at":"2026-05-09T16:16:41Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"99cafecba82d31d45b38b7753fc5ed4b54dc3c30a522023f1bd61277cf64408c","abstract_canon_sha256":"8a981f444072b094ad1688b6e0bcbe4c6bfae8da37ba84d24ae62a67f8434b2a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:00:41.868065Z","signature_b64":"/ARjyAy54JDYO0Obm9ZYtHwEn7VXUTWp1j0ha/rY5u+AnGKul6nzAoSoz7gEvXgrksYyZongBfmy0XlfBjLyAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"13ec189c13caf736f9e0f041aa34fa28e16fe313a39df369e6f1e165f678f80e","last_reissued_at":"2026-05-20T00:00:41.867279Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:00:41.867279Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"ShadowMerge: A Novel Poisoning Attack on Graph-Based Agent Memory via Relation-Channel Conflicts","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"ShadowMerge poisons graph-based agent memory by injecting relations that share the same query-activated anchor and channel as legitimate evidence.","cross_cats":["cs.AI"],"primary_cat":"cs.CR","authors_text":"Lingyun Peng, Shuyu Li, Tiantian Ji, Xinran Liu, Yang Luo, Yong Liu, Zifeng Kang","submitted_at":"2026-05-09T16:16:41Z","abstract_excerpt":"Graph-based agent memory is increasingly used in LLM agents to support structured long-term recall and multi-hop reasoning, but it also creates a new poisoning surface: an attacker can inject a crafted relation into graph memory so that it is later retrieved and influences agent behavior. Existing agent-memory poisoning attacks mainly target flat textual records and are ineffective in graph-based memory because malicious relations often fail to be extracted, merged into the target anchor neighborhood, or retrieved for the victim query.\n  We present SHADOWMERGE, a poisoning attack against graph"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"SHADOWMERGE achieves 93.8% average attack success rate, improving the best baseline by 50.3 absolute points, while having negligible impact on unrelated benign tasks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The graph-memory system will extract, merge into the target anchor neighborhood, and retrieve the poisoned relation for the victim query when it shares the same query-activated anchor and canonicalized relation channel as benign evidence via the AIR pipeline.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"ShadowMerge exploits relation-channel conflicts to poison graph-based agent memory, achieving 93.8% average attack success rate on Mem0 and real-world datasets while bypassing existing defenses.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"ShadowMerge poisons graph-based agent memory by injecting relations that share the same query-activated anchor and channel as legitimate evidence.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"71a62ab75d1b668d9f05cfcf713abd301d691040a175eaee947fb2810fa46b9d"},"source":{"id":"2605.09033","kind":"arxiv","version":3},"verdict":{"id":"203f6772-45e7-42a7-8954-de7a118296b6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T17:34:08.512884Z","strongest_claim":"SHADOWMERGE achieves 93.8% average attack success rate, improving the best baseline by 50.3 absolute points, while having negligible impact on unrelated benign tasks.","one_line_summary":"ShadowMerge exploits relation-channel conflicts to poison graph-based agent memory, achieving 93.8% average attack success rate on Mem0 and real-world datasets while bypassing existing defenses.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The graph-memory system will extract, merge into the target anchor neighborhood, and retrieve the poisoned relation for the victim query when it shares the same query-activated anchor and canonicalized relation channel as benign evidence via the AIR pipeline.","pith_extraction_headline":"ShadowMerge poisons graph-based agent memory by injecting relations that share the same query-activated anchor and channel as legitimate evidence."},"integrity":{"clean":false,"summary":{"advisory":0,"critical":1,"by_detector":{"doi_compliance":{"total":1,"advisory":0,"critical":1,"informational":0}},"informational":0},"endpoint":"/pith/2605.09033/integrity.json","findings":[{"note":"Identifier '10.1109/tkde.2022.3201246' is syntactically valid but the DOI registry (doi.org) returned 404, and Crossref / OpenAlex / internal corpus also have no record. The cited work could not be located through any authoritative source.","detector":"doi_compliance","severity":"critical","ref_index":31,"audited_at":"2026-05-19T10:32:52.510765Z","detected_doi":"10.1109/tkde.2022.3201246","finding_type":"unresolvable_identifier","verdict_class":"cross_source","detected_arxiv_id":null}],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-19T20:38:50.422746Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T13:31:19.305443Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T10:32:52.510765Z","status":"completed","version":"1.0.0","findings_count":1}],"snapshot_sha256":"730e1988477db5891f0e3a5bf459d60d0855b063d73ee968718398565890fcdd"},"references":{"count":57,"sample":[{"doi":"","year":2023,"title":"ReAct: Synergizing Reasoning and Acting in Language Models","work_id":"407a2351-25f1-497d-b611-f77d0292a8e6","ref_index":1,"cited_arxiv_id":"2210.03629","is_internal_anchor":true},{"doi":"","year":2023,"title":"Toolformer: Language Models Can Teach Themselves to Use Tools","work_id":"9bce40c8-cfd7-4983-80e0-c3bd4402322a","ref_index":2,"cited_arxiv_id":"2302.04761","is_internal_anchor":true},{"doi":"","year":2023,"title":"Re- flexion: Language agents with verbal reinforcement learning","work_id":"f1386d00-b17b-4a6f-ab98-78dc28287803","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Voyager: An Open-Ended Embodied Agent with Large Language Models","work_id":"ffe0d207-86cf-4742-a100-e988ac8b9676","ref_index":4,"cited_arxiv_id":"2305.16291","is_internal_anchor":true},{"doi":"10.1145/3586183.3606763","year":2023,"title":"Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , year=","work_id":"b5bf85fe-4fb7-4966-b0b2-9ccf9d3b11b9","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":57,"snapshot_sha256":"06b7cefea1bead7d556bd25bde490e5efa4b46d5e106c2a890a44ce08c996140","internal_anchors":14},"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":"2605.09033","created_at":"2026-05-20T00:00:41.867396+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.09033v3","created_at":"2026-05-20T00:00:41.867396+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.09033","created_at":"2026-05-20T00:00:41.867396+00:00"},{"alias_kind":"pith_short_12","alias_value":"CPWBRHATZL3T","created_at":"2026-05-20T00:00:41.867396+00:00"},{"alias_kind":"pith_short_16","alias_value":"CPWBRHATZL3TN6PA","created_at":"2026-05-20T00:00:41.867396+00:00"},{"alias_kind":"pith_short_8","alias_value":"CPWBRHAT","created_at":"2026-05-20T00:00:41.867396+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CPWBRHATZL3TN6PA6BA2UNH2FD","json":"https://pith.science/pith/CPWBRHATZL3TN6PA6BA2UNH2FD.json","graph_json":"https://pith.science/api/pith-number/CPWBRHATZL3TN6PA6BA2UNH2FD/graph.json","events_json":"https://pith.science/api/pith-number/CPWBRHATZL3TN6PA6BA2UNH2FD/events.json","paper":"https://pith.science/paper/CPWBRHAT"},"agent_actions":{"view_html":"https://pith.science/pith/CPWBRHATZL3TN6PA6BA2UNH2FD","download_json":"https://pith.science/pith/CPWBRHATZL3TN6PA6BA2UNH2FD.json","view_paper":"https://pith.science/paper/CPWBRHAT","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.09033&json=true","fetch_graph":"https://pith.science/api/pith-number/CPWBRHATZL3TN6PA6BA2UNH2FD/graph.json","fetch_events":"https://pith.science/api/pith-number/CPWBRHATZL3TN6PA6BA2UNH2FD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CPWBRHATZL3TN6PA6BA2UNH2FD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CPWBRHATZL3TN6PA6BA2UNH2FD/action/storage_attestation","attest_author":"https://pith.science/pith/CPWBRHATZL3TN6PA6BA2UNH2FD/action/author_attestation","sign_citation":"https://pith.science/pith/CPWBRHATZL3TN6PA6BA2UNH2FD/action/citation_signature","submit_replication":"https://pith.science/pith/CPWBRHATZL3TN6PA6BA2UNH2FD/action/replication_record"}},"created_at":"2026-05-20T00:00:41.867396+00:00","updated_at":"2026-05-20T00:00:41.867396+00:00"}