{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2023:HVCG4CD3HCUZZSWBBEQJSTTNJI","short_pith_number":"pith:HVCG4CD3","canonical_record":{"source":{"id":"2308.14132","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2023-08-27T15:20:06Z","cross_cats_sorted":["cs.AI","cs.CR","cs.LG"],"title_canon_sha256":"f4426ad31296763e835ab54ae2f81682dc2297e18b9ace3e38a91745fb9b4ca0","abstract_canon_sha256":"ec4a9746d99896900950ab9f3086052ada3108e7580cbd70a19fafc096b18af1"},"schema_version":"1.0"},"canonical_sha256":"3d446e087b38a99ccac10920994e6d4a1c6dbdc4f1862e05219bdd4860492e3b","source":{"kind":"arxiv","id":"2308.14132","version":3},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2308.14132","created_at":"2026-05-17T23:38:52Z"},{"alias_kind":"arxiv_version","alias_value":"2308.14132v3","created_at":"2026-05-17T23:38:52Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2308.14132","created_at":"2026-05-17T23:38:52Z"},{"alias_kind":"pith_short_12","alias_value":"HVCG4CD3HCUZ","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_16","alias_value":"HVCG4CD3HCUZZSWB","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_8","alias_value":"HVCG4CD3","created_at":"2026-05-18T12:33:33Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2023:HVCG4CD3HCUZZSWBBEQJSTTNJI","target":"record","payload":{"canonical_record":{"source":{"id":"2308.14132","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2023-08-27T15:20:06Z","cross_cats_sorted":["cs.AI","cs.CR","cs.LG"],"title_canon_sha256":"f4426ad31296763e835ab54ae2f81682dc2297e18b9ace3e38a91745fb9b4ca0","abstract_canon_sha256":"ec4a9746d99896900950ab9f3086052ada3108e7580cbd70a19fafc096b18af1"},"schema_version":"1.0"},"canonical_sha256":"3d446e087b38a99ccac10920994e6d4a1c6dbdc4f1862e05219bdd4860492e3b","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:52.385359Z","signature_b64":"fwSVI2MGpHL3cr2mpO6pAtseVWKtLa6/YtXJwfuwqKKuSwG4Qqs9/OVA4m0bg40GZjuvpXCHRLLdsWrpDyzNAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3d446e087b38a99ccac10920994e6d4a1c6dbdc4f1862e05219bdd4860492e3b","last_reissued_at":"2026-05-17T23:38:52.384790Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:52.384790Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2308.14132","source_version":3,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:52Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"/DygCEYOfNcECfeJAue/laIl/cmX4huL52JvO9EVcHGTv4uZl1B+/xZ7itcYWJ1jyUTJnBcSAo+h4aWO8Gd3Dw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-02T08:42:51.363819Z"},"content_sha256":"1496cf1d9cda5e717693c168917a3a235becc3f51491244a7c9d6069a8d72683","schema_version":"1.0","event_id":"sha256:1496cf1d9cda5e717693c168917a3a235becc3f51491244a7c9d6069a8d72683"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2023:HVCG4CD3HCUZZSWBBEQJSTTNJI","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Detecting Language Model Attacks with Perplexity","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Adversarial jailbreak suffixes produce high perplexity under GPT-2, allowing a classifier on perplexity and length to catch most attacks.","cross_cats":["cs.AI","cs.CR","cs.LG"],"primary_cat":"cs.CL","authors_text":"Gabriel Alon, Michael Kamfonas","submitted_at":"2023-08-27T15:20:06Z","abstract_excerpt":"A novel hack involving Large Language Models (LLMs) has emerged, exploiting adversarial suffixes to deceive models into generating perilous responses. Such jailbreaks can trick LLMs into providing intricate instructions to a malicious user for creating explosives, orchestrating a bank heist, or facilitating the creation of offensive content. By evaluating the perplexity of queries with adversarial suffixes using an open-source LLM (GPT-2), we found that they have exceedingly high perplexity values. As we explored a broad range of regular (non-adversarial) prompt varieties, we concluded that fa"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"By evaluating the perplexity of queries with adversarial suffixes using an open-source LLM (GPT-2), we found that they have exceedingly high perplexity values. [...] A Light-GBM trained on perplexity and token length resolved the false positives and correctly detected most adversarial attacks in the test set.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the distribution of regular (non-adversarial) prompts used to measure false positives is representative of real-world usage and that future attackers will not adapt suffixes to also produce low perplexity under GPT-2.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Jailbreak prompts with adversarial suffixes have high GPT-2 perplexity, and a LightGBM model on perplexity and length detects most attacks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Adversarial jailbreak suffixes produce high perplexity under GPT-2, allowing a classifier on perplexity and length to catch most attacks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e5a6f0fd51229d0f1a136796718c5ecc2ee6b912899a58e60e629baa18324bd4"},"source":{"id":"2308.14132","kind":"arxiv","version":3},"verdict":{"id":"e55be181-2a61-4cb2-8954-7bca598aa37f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T13:55:22.051094Z","strongest_claim":"By evaluating the perplexity of queries with adversarial suffixes using an open-source LLM (GPT-2), we found that they have exceedingly high perplexity values. [...] A Light-GBM trained on perplexity and token length resolved the false positives and correctly detected most adversarial attacks in the test set.","one_line_summary":"Jailbreak prompts with adversarial suffixes have high GPT-2 perplexity, and a LightGBM model on perplexity and length detects most attacks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the distribution of regular (non-adversarial) prompts used to measure false positives is representative of real-world usage and that future attackers will not adapt suffixes to also produce low perplexity under GPT-2.","pith_extraction_headline":"Adversarial jailbreak suffixes produce high perplexity under GPT-2, allowing a classifier on perplexity and length to catch most attacks."},"references":{"count":83,"sample":[{"doi":"","year":2022,"title":"Training a helpful and harmless assistant with reinforcement learning from human feedback, 2022","work_id":"6b9cf002-ab59-4c61-ae20-2d2f7b0eecaf","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Boolq: Exploring the surprising difficulty of natural yes/no questions","work_id":"95712603-7f1e-44dd-825a-da30fd36d3aa","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Certified adversarial robustness via randomized smoothing","work_id":"d07eec87-9b6a-4a0e-b8f8-aee82051d662","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2012,"title":"Monitor alarm fatigue: an integrative review","work_id":"95443dd5-754f-4923-a4fa-453decbf764d","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Improving alignment of dialogue agents via targeted human judgments, 2022","work_id":"279da37f-eca7-468c-9da9-7d42921a9b93","ref_index":6,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":83,"snapshot_sha256":"09a310974125e1a6db0726b75ef3260df968cc1eb1c26603f3445e630ce39dbf","internal_anchors":4},"formal_canon":{"evidence_count":2,"snapshot_sha256":"c87afabb31051e2af7cd34f5765435442bab6e4a22119dc4d57870ea2efe33f8"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"e55be181-2a61-4cb2-8954-7bca598aa37f"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:52Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"1hBzjF92ySo1wfaLle0RwOYL1P3E4KY5Fa/y4Sk2M0I3o6vXtamco6oA1o4uKQaSkKcsDvaZfT6+IDaoz27KAw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-02T08:42:51.364364Z"},"content_sha256":"5c7c7192993a8f2c69b23245a2d2396b03bda91a866b4248c5f76dadb9f04355","schema_version":"1.0","event_id":"sha256:5c7c7192993a8f2c69b23245a2d2396b03bda91a866b4248c5f76dadb9f04355"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/HVCG4CD3HCUZZSWBBEQJSTTNJI/bundle.json","state_url":"https://pith.science/pith/HVCG4CD3HCUZZSWBBEQJSTTNJI/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/HVCG4CD3HCUZZSWBBEQJSTTNJI/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-06-02T08:42:51Z","links":{"resolver":"https://pith.science/pith/HVCG4CD3HCUZZSWBBEQJSTTNJI","bundle":"https://pith.science/pith/HVCG4CD3HCUZZSWBBEQJSTTNJI/bundle.json","state":"https://pith.science/pith/HVCG4CD3HCUZZSWBBEQJSTTNJI/state.json","well_known_bundle":"https://pith.science/.well-known/pith/HVCG4CD3HCUZZSWBBEQJSTTNJI/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2023:HVCG4CD3HCUZZSWBBEQJSTTNJI","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"ec4a9746d99896900950ab9f3086052ada3108e7580cbd70a19fafc096b18af1","cross_cats_sorted":["cs.AI","cs.CR","cs.LG"],"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2023-08-27T15:20:06Z","title_canon_sha256":"f4426ad31296763e835ab54ae2f81682dc2297e18b9ace3e38a91745fb9b4ca0"},"schema_version":"1.0","source":{"id":"2308.14132","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2308.14132","created_at":"2026-05-17T23:38:52Z"},{"alias_kind":"arxiv_version","alias_value":"2308.14132v3","created_at":"2026-05-17T23:38:52Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2308.14132","created_at":"2026-05-17T23:38:52Z"},{"alias_kind":"pith_short_12","alias_value":"HVCG4CD3HCUZ","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_16","alias_value":"HVCG4CD3HCUZZSWB","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_8","alias_value":"HVCG4CD3","created_at":"2026-05-18T12:33:33Z"}],"graph_snapshots":[{"event_id":"sha256:5c7c7192993a8f2c69b23245a2d2396b03bda91a866b4248c5f76dadb9f04355","target":"graph","created_at":"2026-05-17T23:38:52Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"By evaluating the perplexity of queries with adversarial suffixes using an open-source LLM (GPT-2), we found that they have exceedingly high perplexity values. [...] A Light-GBM trained on perplexity and token length resolved the false positives and correctly detected most adversarial attacks in the test set."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the distribution of regular (non-adversarial) prompts used to measure false positives is representative of real-world usage and that future attackers will not adapt suffixes to also produce low perplexity under GPT-2."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Jailbreak prompts with adversarial suffixes have high GPT-2 perplexity, and a LightGBM model on perplexity and length detects most attacks."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Adversarial jailbreak suffixes produce high perplexity under GPT-2, allowing a classifier on perplexity and length to catch most attacks."}],"snapshot_sha256":"e5a6f0fd51229d0f1a136796718c5ecc2ee6b912899a58e60e629baa18324bd4"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"c87afabb31051e2af7cd34f5765435442bab6e4a22119dc4d57870ea2efe33f8"},"paper":{"abstract_excerpt":"A novel hack involving Large Language Models (LLMs) has emerged, exploiting adversarial suffixes to deceive models into generating perilous responses. Such jailbreaks can trick LLMs into providing intricate instructions to a malicious user for creating explosives, orchestrating a bank heist, or facilitating the creation of offensive content. By evaluating the perplexity of queries with adversarial suffixes using an open-source LLM (GPT-2), we found that they have exceedingly high perplexity values. As we explored a broad range of regular (non-adversarial) prompt varieties, we concluded that fa","authors_text":"Gabriel Alon, Michael Kamfonas","cross_cats":["cs.AI","cs.CR","cs.LG"],"headline":"Adversarial jailbreak suffixes produce high perplexity under GPT-2, allowing a classifier on perplexity and length to catch most attacks.","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2023-08-27T15:20:06Z","title":"Detecting Language Model Attacks with Perplexity"},"references":{"count":83,"internal_anchors":4,"resolved_work":83,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Training a helpful and harmless assistant with reinforcement learning from human feedback, 2022","work_id":"6b9cf002-ab59-4c61-ae20-2d2f7b0eecaf","year":2022},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"Boolq: Exploring the surprising difficulty of natural yes/no questions","work_id":"95712603-7f1e-44dd-825a-da30fd36d3aa","year":2019},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Certified adversarial robustness via randomized smoothing","work_id":"d07eec87-9b6a-4a0e-b8f8-aee82051d662","year":2019},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Monitor alarm fatigue: an integrative review","work_id":"95443dd5-754f-4923-a4fa-453decbf764d","year":2012},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":6,"title":"Improving alignment of dialogue agents via targeted human judgments, 2022","work_id":"279da37f-eca7-468c-9da9-7d42921a9b93","year":2022}],"snapshot_sha256":"09a310974125e1a6db0726b75ef3260df968cc1eb1c26603f3445e630ce39dbf"},"source":{"id":"2308.14132","kind":"arxiv","version":3},"verdict":{"created_at":"2026-05-15T13:55:22.051094Z","id":"e55be181-2a61-4cb2-8954-7bca598aa37f","model_set":{"reader":"grok-4.3"},"one_line_summary":"Jailbreak prompts with adversarial suffixes have high GPT-2 perplexity, and a LightGBM model on perplexity and length detects most attacks.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Adversarial jailbreak suffixes produce high perplexity under GPT-2, allowing a classifier on perplexity and length to catch most attacks.","strongest_claim":"By evaluating the perplexity of queries with adversarial suffixes using an open-source LLM (GPT-2), we found that they have exceedingly high perplexity values. [...] A Light-GBM trained on perplexity and token length resolved the false positives and correctly detected most adversarial attacks in the test set.","weakest_assumption":"That the distribution of regular (non-adversarial) prompts used to measure false positives is representative of real-world usage and that future attackers will not adapt suffixes to also produce low perplexity under GPT-2."}},"verdict_id":"e55be181-2a61-4cb2-8954-7bca598aa37f"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:1496cf1d9cda5e717693c168917a3a235becc3f51491244a7c9d6069a8d72683","target":"record","created_at":"2026-05-17T23:38:52Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"ec4a9746d99896900950ab9f3086052ada3108e7580cbd70a19fafc096b18af1","cross_cats_sorted":["cs.AI","cs.CR","cs.LG"],"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2023-08-27T15:20:06Z","title_canon_sha256":"f4426ad31296763e835ab54ae2f81682dc2297e18b9ace3e38a91745fb9b4ca0"},"schema_version":"1.0","source":{"id":"2308.14132","kind":"arxiv","version":3}},"canonical_sha256":"3d446e087b38a99ccac10920994e6d4a1c6dbdc4f1862e05219bdd4860492e3b","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"3d446e087b38a99ccac10920994e6d4a1c6dbdc4f1862e05219bdd4860492e3b","first_computed_at":"2026-05-17T23:38:52.384790Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:52.384790Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"fwSVI2MGpHL3cr2mpO6pAtseVWKtLa6/YtXJwfuwqKKuSwG4Qqs9/OVA4m0bg40GZjuvpXCHRLLdsWrpDyzNAw==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:52.385359Z","signed_message":"canonical_sha256_bytes"},"source_id":"2308.14132","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:1496cf1d9cda5e717693c168917a3a235becc3f51491244a7c9d6069a8d72683","sha256:5c7c7192993a8f2c69b23245a2d2396b03bda91a866b4248c5f76dadb9f04355"],"state_sha256":"04e57718e3f31c46779283ca61530fdb254b13b205f99df37bfe6531020019db"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"fwuVg0hlcD2O1PWmaNKdAMfKvyJQasvpz4gXjZpjHsLa35IByLIQcf/vxrlKJWE7gkcqCq/JMvdA2icEJuaWDw==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-02T08:42:51.366832Z","bundle_sha256":"6141b0be7b97f2f6fa2431c094394c9ac7ca21e22fb63de64d13cf3c78b5279c"}}