遇见数据集

AgentPROV corpora: AgentDojo-format studies I (MIT)

收藏
Zenodo2026-10-30 更新2026-11-05 收录
官方服务:

资源简介:

Execution-provenance corpora built by AgentPROV v0.1.0 from third-party AgentDojo-format run logs published under the MIT licence (DRIFT and its derived studies, CrAIBench, Glitch, Kavach and others): each execution as an AgentPROV Execution Record, semantic annotations and W3C PROV bundles, validated and hash-chained. This deposit draft holds 8 AgentPROV corpus version(s), prepared offline by agentprov package --deposit for the release date 2026-10-30. Nothing has been uploaded or published. Corpus ceferisbarov-agentdojo 0.1.0: ceferisbarov/agentdojo reruns: AgentDojo logs of optimised injection strings over 61 arms (ceferisbarov/agentdojo). 18438 execution records, of which 5 are outcome-only records. Licence class: MIT. Content digest sha256:1c24b68a83174c3632d10fb2d9af2b5300293120693a5ac6a1fa3add3cc61936; root digest sha256:f2ec03ce15f0bfde06c043739e72bffd08b33cb7b6c08a6ea0f2d6727502df8f. Corpus craibench-agentdojo 0.1.0: CrAIBench: AgentDojo v0.k11 logs of two new crypto suites with memory and prompt injection (AthaSSiN/CrAIBench). 1163 execution records, of which 0 are outcome-only records. Licence class: MIT. Content digest sha256:9a304f5dd60bdb9d04e6eb5d12272e937c6a76e6335a163033f202e325bc7fb2; root digest sha256:75627cd2bb9c7e329cbf47d2f7beef2ac5a36c805c057d63e3cb5dd965881548. Corpus dca-agentdojo 0.1.0: DCA AgentDojo artefact v1: AgentDojo 0.1.35 logs of gpt-4o-mini, GPT-OSS-20B and Qwen models under DCA, CaMeL and DRIFT (Figshare 32657856). 16211 execution records, of which 0 are outcome-only records. Licence class: MIT. Content digest sha256:11270f86d9e2c044d12f889bd17c0b6749c8b5103c872b1e24ebce5df33d6a12; root digest sha256:2d8592f352713aef9e75a9f71b8e37af036f5ff9421d88175985ed55c90f6bd2. Corpus drift-head 0.1.0: DRIFT (SaFo-Lab/DRIFT) run logs: AgentDojo v1.2 and AgentDyn suites under the DRIFT defence, gpt-4o-mini. 1564 execution records, of which 0 are outcome-only records. Licence class: MIT. Content digest sha256:b38dc5d485c637ba8c3fdd9ead29e53d34092e71648b9d92c9e156bef3e43fb3; root digest sha256:bcba237ef6281bf7f005033de6f0329faa22929cc6143df34f64a877e60097c0. Corpus drift1-agentdojo 0.1.0: drift-1 (CGboom-git/drift-1): DRIFT-format logs of gpt-4o-mini, including a PACT-DRIFT arm. 626 execution records, of which 0 are outcome-only records. Licence class: MIT. Content digest sha256:1d825038f6873c09d4362f87e1f39fbec9ecbd1b86cf26ba2213f4263fdf5d08; root digest sha256:ad79a9e1f66fdb31f60ec4eb2cf817b91c3a2455233ee663dff1475f5caafe57. Corpus glitch-agentdojo 0.1.0: Glitch on AgentDojo: AgentDojo 0.1.35 logs of an external tool-call gateway (zanethomas1430-code/Glitch). 1842 execution records, of which 0 are outcome-only records. Licence class: MIT. Content digest sha256:80d94a9b30b1bb43e4fc40fe323fcba0ac3d3485cc23275e766e36d18d7ed732; root digest sha256:af726e4aa4b9b8031693aabe40f5b109cd551f22c154a94963a02314c0536626. Corpus hilarytorn-trajectory-poisoning 0.1.0: In-context trajectory poisoning: AgentDojo 0.1.35 slack logs of gpt-4o (HilaryTorn/in-context-trajectory-poisoning). 990 execution records, of which 0 are outcome-only records. Licence class: MIT. Content digest sha256:0cd16834b7f55b6373bfdccea0fba0a4c81a79f3b9e0104f89c139d32f9aebb3; root digest sha256:5f5e09eb75310216cc5646b3e3dbba76c7ec15abb30b0c160c66ebf987f99194. Corpus kavach-agentdojo 0.1.0: Kavach: AgentDojo 0.1.35 logs of a runtime tool-call defence on local Ollama models (Ishani018/Kavach, three branch pins). 2619 execution records, of which 0 are outcome-only records. Licence class: MIT. Content digest sha256:a8b6f7ed48ba9b9b21e34abf13ba01cf425ca6e97ba925047cde94b937935832; root digest sha256:110db3c05e69f98291c8a09aab59a59da0468bf297192114f0c3cd3daab4dc62. Counts are computed from each corpus manifest at sealing. Each DATASHEET states what the records do and do not establish; a statement about the data is scoped as it is there. To verify: compare each archive's sha256 or md5 with DEPOSIT-MANIFEST.json, run agentprov verify on the archives (archive mode, no extraction; H.5.5), and compare each root digest with the one published outside the deposit (the corpus tag annotation and the paper).

提供机构:
Zenodo
创建时间:
2026-10-30
二维码
社区交流群
二维码
科研交流群
商业服务