{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:Q6TUQA2AIJLE2IDHIO3MI6QHLO","short_pith_number":"pith:Q6TUQA2A","schema_version":"1.0","canonical_sha256":"87a748034042564d206743b6c47a075b97ae7ed330f23b45e418c4b3d1d99da1","source":{"kind":"arxiv","id":"2502.07987","version":3},"attestation_state":"computed","paper":{"title":"Universal Adversarial Attack on Aligned Multimodal LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Andrey Kuznetsov, Anton Razzhigaev, Matvey Mikhalchuk, Nikita Kurdiukov, Polina Druzhinina, Temurbek Rahmatullaev","submitted_at":"2025-02-11T22:07:47Z","abstract_excerpt":"We propose a universal adversarial attack on multimodal Large Language Models (LLMs) that leverages a single optimized image to override alignment safeguards across diverse queries and even multiple models. By backpropagating through the vision encoder and language head, we craft a synthetic image that forces the model to respond with a targeted phrase (e.g., \"Sure, here it is\") or otherwise unsafe content -- even for harmful prompts. In experiments on the SafeBench and MM-SafetyBench benchmarks, our method achieves higher attack success rates than existing baselines, including text-only unive"},"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":"2502.07987","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2025-02-11T22:07:47Z","cross_cats_sorted":[],"title_canon_sha256":"861a42a517fe43e5dac7009f0ec42bb3e8e419aa7414963e483f0d64a49e0b9f","abstract_canon_sha256":"61a9691a79d172271b39ad2ad6db02720fad7f65fe526405daf662d372d52c87"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:16:06.089356Z","signature_b64":"DSX6hmPBVdWzhMn3QqLsGtnyYt43nhJRQr+2fIfjtKGLlRH5Y3p5jhFomJQYHarKSWQ0NxmwnOtz7sovOeQ1BA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"87a748034042564d206743b6c47a075b97ae7ed330f23b45e418c4b3d1d99da1","last_reissued_at":"2026-07-05T11:16:06.088839Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:16:06.088839Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Universal Adversarial Attack on Aligned Multimodal LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Andrey Kuznetsov, Anton Razzhigaev, Matvey Mikhalchuk, Nikita Kurdiukov, Polina Druzhinina, Temurbek Rahmatullaev","submitted_at":"2025-02-11T22:07:47Z","abstract_excerpt":"We propose a universal adversarial attack on multimodal Large Language Models (LLMs) that leverages a single optimized image to override alignment safeguards across diverse queries and even multiple models. By backpropagating through the vision encoder and language head, we craft a synthetic image that forces the model to respond with a targeted phrase (e.g., \"Sure, here it is\") or otherwise unsafe content -- even for harmful prompts. In experiments on the SafeBench and MM-SafetyBench benchmarks, our method achieves higher attack success rates than existing baselines, including text-only unive"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.07987","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/2502.07987/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":"2502.07987","created_at":"2026-07-05T11:16:06.088911+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.07987v3","created_at":"2026-07-05T11:16:06.088911+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.07987","created_at":"2026-07-05T11:16:06.088911+00:00"},{"alias_kind":"pith_short_12","alias_value":"Q6TUQA2AIJLE","created_at":"2026-07-05T11:16:06.088911+00:00"},{"alias_kind":"pith_short_16","alias_value":"Q6TUQA2AIJLE2IDH","created_at":"2026-07-05T11:16:06.088911+00:00"},{"alias_kind":"pith_short_8","alias_value":"Q6TUQA2A","created_at":"2026-07-05T11:16:06.088911+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2511.21893","citing_title":"Breaking the Illusion: Consensus-Based Generative Mitigation of Adversarial Illusions in Multi-Modal Embeddings","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2605.01449","citing_title":"VisInject: Disruption != Injection -- A Dual-Dimension Evaluation of Universal Adversarial Attacks on Vision-Language Models","ref_index":25,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/Q6TUQA2AIJLE2IDHIO3MI6QHLO","json":"https://pith.science/pith/Q6TUQA2AIJLE2IDHIO3MI6QHLO.json","graph_json":"https://pith.science/api/pith-number/Q6TUQA2AIJLE2IDHIO3MI6QHLO/graph.json","events_json":"https://pith.science/api/pith-number/Q6TUQA2AIJLE2IDHIO3MI6QHLO/events.json","paper":"https://pith.science/paper/Q6TUQA2A"},"agent_actions":{"view_html":"https://pith.science/pith/Q6TUQA2AIJLE2IDHIO3MI6QHLO","download_json":"https://pith.science/pith/Q6TUQA2AIJLE2IDHIO3MI6QHLO.json","view_paper":"https://pith.science/paper/Q6TUQA2A","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.07987&json=true","fetch_graph":"https://pith.science/api/pith-number/Q6TUQA2AIJLE2IDHIO3MI6QHLO/graph.json","fetch_events":"https://pith.science/api/pith-number/Q6TUQA2AIJLE2IDHIO3MI6QHLO/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/Q6TUQA2AIJLE2IDHIO3MI6QHLO/action/timestamp_anchor","attest_storage":"https://pith.science/pith/Q6TUQA2AIJLE2IDHIO3MI6QHLO/action/storage_attestation","attest_author":"https://pith.science/pith/Q6TUQA2AIJLE2IDHIO3MI6QHLO/action/author_attestation","sign_citation":"https://pith.science/pith/Q6TUQA2AIJLE2IDHIO3MI6QHLO/action/citation_signature","submit_replication":"https://pith.science/pith/Q6TUQA2AIJLE2IDHIO3MI6QHLO/action/replication_record"}},"created_at":"2026-07-05T11:16:06.088911+00:00","updated_at":"2026-07-05T11:16:06.088911+00:00"}