{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:NG5QIF47MSEUVG4FYB34NER6HH","short_pith_number":"pith:NG5QIF47","schema_version":"1.0","canonical_sha256":"69bb04179f64894a9b85c077c6923e39e762c60311d5959e8c52ada9ceb41723","source":{"kind":"arxiv","id":"2407.03232","version":1},"attestation_state":"computed","paper":{"title":"Single Character Perturbations Break LLM Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Hannah Brown, Kenji Kawaguchi, Leon Lin, Michael Shieh","submitted_at":"2024-07-03T16:03:10Z","abstract_excerpt":"When LLMs are deployed in sensitive, human-facing settings, it is crucial that they do not output unsafe, biased, or privacy-violating outputs. For this reason, models are both trained and instructed to refuse to answer unsafe prompts such as \"Tell me how to build a bomb.\" We find that, despite these safeguards, it is possible to break model defenses simply by appending a space to the end of a model's input. In a study of eight open-source models, we demonstrate that this acts as a strong enough attack to cause the majority of models to generate harmful outputs with very high success rates. We"},"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":"2407.03232","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-07-03T16:03:10Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"62435e10220147c0ed9baa0fa052c2194b927d226a02f640a549bbf6a33ce457","abstract_canon_sha256":"4b404f5320196e6302122d0dfa40aeee1ba87c11a573e2cce0f87f9e8c4e252e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:39:48.054512Z","signature_b64":"55BgDEwI71s1h9qX38o7vKwevqNvXf0HljeX4verz5lho/p4vDwcfsT5/taG7w77lZ6/XKxcwUBMtaC+btpGAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"69bb04179f64894a9b85c077c6923e39e762c60311d5959e8c52ada9ceb41723","last_reissued_at":"2026-07-05T08:39:48.054077Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:39:48.054077Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Single Character Perturbations Break LLM Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Hannah Brown, Kenji Kawaguchi, Leon Lin, Michael Shieh","submitted_at":"2024-07-03T16:03:10Z","abstract_excerpt":"When LLMs are deployed in sensitive, human-facing settings, it is crucial that they do not output unsafe, biased, or privacy-violating outputs. For this reason, models are both trained and instructed to refuse to answer unsafe prompts such as \"Tell me how to build a bomb.\" We find that, despite these safeguards, it is possible to break model defenses simply by appending a space to the end of a model's input. In a study of eight open-source models, we demonstrate that this acts as a strong enough attack to cause the majority of models to generate harmful outputs with very high success rates. We"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2407.03232","kind":"arxiv","version":1},"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/2407.03232/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":"2407.03232","created_at":"2026-07-05T08:39:48.054130+00:00"},{"alias_kind":"arxiv_version","alias_value":"2407.03232v1","created_at":"2026-07-05T08:39:48.054130+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2407.03232","created_at":"2026-07-05T08:39:48.054130+00:00"},{"alias_kind":"pith_short_12","alias_value":"NG5QIF47MSEU","created_at":"2026-07-05T08:39:48.054130+00:00"},{"alias_kind":"pith_short_16","alias_value":"NG5QIF47MSEUVG4F","created_at":"2026-07-05T08:39:48.054130+00:00"},{"alias_kind":"pith_short_8","alias_value":"NG5QIF47","created_at":"2026-07-05T08:39:48.054130+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2501.18626","citing_title":"The TIP of the Iceberg: Revealing a Hidden Class of Task-in-Prompt Adversarial Attacks on LLMs","ref_index":11,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/NG5QIF47MSEUVG4FYB34NER6HH","json":"https://pith.science/pith/NG5QIF47MSEUVG4FYB34NER6HH.json","graph_json":"https://pith.science/api/pith-number/NG5QIF47MSEUVG4FYB34NER6HH/graph.json","events_json":"https://pith.science/api/pith-number/NG5QIF47MSEUVG4FYB34NER6HH/events.json","paper":"https://pith.science/paper/NG5QIF47"},"agent_actions":{"view_html":"https://pith.science/pith/NG5QIF47MSEUVG4FYB34NER6HH","download_json":"https://pith.science/pith/NG5QIF47MSEUVG4FYB34NER6HH.json","view_paper":"https://pith.science/paper/NG5QIF47","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2407.03232&json=true","fetch_graph":"https://pith.science/api/pith-number/NG5QIF47MSEUVG4FYB34NER6HH/graph.json","fetch_events":"https://pith.science/api/pith-number/NG5QIF47MSEUVG4FYB34NER6HH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/NG5QIF47MSEUVG4FYB34NER6HH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/NG5QIF47MSEUVG4FYB34NER6HH/action/storage_attestation","attest_author":"https://pith.science/pith/NG5QIF47MSEUVG4FYB34NER6HH/action/author_attestation","sign_citation":"https://pith.science/pith/NG5QIF47MSEUVG4FYB34NER6HH/action/citation_signature","submit_replication":"https://pith.science/pith/NG5QIF47MSEUVG4FYB34NER6HH/action/replication_record"}},"created_at":"2026-07-05T08:39:48.054130+00:00","updated_at":"2026-07-05T08:39:48.054130+00:00"}