ADR/docs · adr10/ADR

GGitHubfeat(detection): add privacy-bounded run provenance manifest (#126)

文件	最后提交记录	最后更新时间
BASELINE_REPLICATION.md	Initial public release: ADR Sensor, Detector, and ADR-Bench Research artifacts for "ADR: An Agentic Detection System for Enterprise Agentic AI Security" (MLSys 2026). Includes: - ADR Sensor (Sensor/): telemetry capture for AI agent / MCP activity - ADR Detector (Detection/): dual-agent detection framework - ADR-Bench: 133 MCP servers, 303 benchmark tasks, four detector baselines - Reproducibility guide, open-source review, baseline replication docs - Isolation/non-production notice (research artifact, pinned deps carry known CVEs acceptable under isolated threat model) Naming aligned with the paper: system = ADR, benchmark = ADR-Bench. The internal codebase previously used "ADS" / "ADS-Bench" (pre-rename); this release renames all user-facing and most internal identifiers (package name adr-benchmark, CLI --detector adr / --benchmark adr_bench, module paths guardrail/adr_agent/adr_baseline, ADRBaseline class, adr_framework config key, threat IDs ADR.T00XX, display strings) to match the paper. A few internal-only identifiers remain (ADSConfig, load_ads_tasks, the ads_reasoning_workspace runtime dir name); these are not user-facing and do not affect imports, CLI, or config wiring. Historical captured-session content inside the packed benchmark JSONL (paths, the benchmark_name field) is preserved as-is since it is non-functional artifact data. Benchmark fixture documents: two real arXiv papers licensed under CC-BY-4.0 are bundled as sample documents for benign document-conversion tasks (see NOTICE.md Components 6 and 7 for attributions). Paper: https://arxiv.org/abs/2605.17380	2 个月前
OPEN_SOURCE_REVIEW.md	Initial public release: ADR Sensor, Detector, and ADR-Bench Research artifacts for "ADR: An Agentic Detection System for Enterprise Agentic AI Security" (MLSys 2026). Includes: - ADR Sensor (Sensor/): telemetry capture for AI agent / MCP activity - ADR Detector (Detection/): dual-agent detection framework - ADR-Bench: 133 MCP servers, 303 benchmark tasks, four detector baselines - Reproducibility guide, open-source review, baseline replication docs - Isolation/non-production notice (research artifact, pinned deps carry known CVEs acceptable under isolated threat model) Naming aligned with the paper: system = ADR, benchmark = ADR-Bench. The internal codebase previously used "ADS" / "ADS-Bench" (pre-rename); this release renames all user-facing and most internal identifiers (package name adr-benchmark, CLI --detector adr / --benchmark adr_bench, module paths guardrail/adr_agent/adr_baseline, ADRBaseline class, adr_framework config key, threat IDs ADR.T00XX, display strings) to match the paper. A few internal-only identifiers remain (ADSConfig, load_ads_tasks, the ads_reasoning_workspace runtime dir name); these are not user-facing and do not affect imports, CLI, or config wiring. Historical captured-session content inside the packed benchmark JSONL (paths, the benchmark_name field) is preserved as-is since it is non-functional artifact data. Benchmark fixture documents: two real arXiv papers licensed under CC-BY-4.0 are bundled as sample documents for benign document-conversion tasks (see NOTICE.md Components 6 and 7 for attributions). Paper: https://arxiv.org/abs/2605.17380	2 个月前
RELEASING.md	release(sensor): add package publishing workflow	2 个月前
REPRODUCIBILITY.md	feat(detection): add privacy-bounded run provenance manifest (#126) * feat(detection): collect run provenance manifest in detector results Added run_manifest collection to benchmark analysis. * feat(detection): add privacy-bounded run manifest collector This file implements a provenance collector for detector analysis output, including SHA256 hashing for files and Git metadata. * docs(detection): document run manifest in detector results Added details about detector results and provenance. * test(detection): assert run manifest in detector results Added assertions for run_manifest in test_main_detector.py. * test(detection): add run manifest unit tests This file contains tests for the privacy-safe detector run provenance, including functionality for hashing conversations, handling ground truth files, and managing git commands. * docs: describe run manifest in reproducibility guide Added details about run_manifest for run provenance and its contents. * fix(detection): bind run provenance to consumed inputs Include run_manifest in the installed wheel. Capture input digests while loading configuration, tasks, ground truth, and conversations, and capture source metadata before analysis. Reuse one task snapshot for labels and MCP definitions. Add regression tests for packaging, working-directory inputs, mutations during analysis, missing metadata, and analyzer reuse. Clarify snapshot behavior in the documentation. --------- Co-authored-by: Baris Ozbas <bozbas@uber.com>	1 个月前
adr-mlsys-2026-slides.pdf	Add ADR paper and MLSys slides	2 个月前
adr-paper.pdf	Add ADR paper and MLSys slides	2 个月前