Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
pdf
pdf

Agent Security Event Reports (Sixu AI · Moyan Audit)

The single-event report series of Sixu AI · Moyan Audit. Each report takes one publicly reported AI / autonomous-agent security event, adjudicates it against the V19 audit rule set, and records a coverage verdict together with its evidence basis. Reports are published per event, on the day the event is triaged wherever possible, and are never silently rewritten after publication — corrections are appended as errata.

Evidence caliber — read this before citing a report

  • Documentary evidence only. Reports rest on public sources (vendor advisories, CVE/GHSA records, security press). They contain no live-run data: no claim in any report rests on executing an audited system end-to-end.
  • Coverage verdicts are static, rule-level verification results — which rules exist, when each was introduced relative to the event's public date, and whether the rule set's evidence model can express the case. A verdict is not a runtime result and must not be cited as one.
  • Review level. Each report records its review level. L2 denotes a completed non-author review pass within the Sixu AI · Moyan Audit team, bound to the exact file hash it reviewed; a verdict or file that changes after review is re-reviewed before republication.
  • Verdict vocabulary (shared with the exam bank): prediction hit (a rule predating the event's first public date covers it) · post-hoc coverage (a rule written after the event covers it) · not covered · unverified. Retrospective coverage is never reported as a prediction.

Relation to the exam bank

Each report cites the exam item it adjudicates (EXAM-YYYYMMDD-NN) from the companion dataset sixujinzhu/agent-security-exam-bank. The exam bank is the item ledger (one line per event); this dataset holds the full per-event reports behind those lines.

Directory convention

One folder per report under reports/<TB-ID>/, containing that report's files (Markdown source and PDF rendering; a -EN / -ZH suffix marks the language where more than one language edition exists). New reports are appended under their own TB-ID folder; this convention applies to all future reports in the series.

Reports index

TB-ID Exam item Event Coverage verdict Files
TB-014 EXAM-20261119-01 Unauthenticated stored XSS in AI Puffer via AI-generated comment replies (CVE-2026-130121) Not covered (static verification) EN Markdown + EN PDF

TB-014 is also published at https://sixu-ai.net.cn/reports/tb-014.html and in the GitHub mirror https://github.com/Liuyanfeng1234/sixu-research (Markdown only).

Maturity label

Working report series. Each report is a working report under the maintainer's maturity labels (dataset → working report → formal report); a label is only ever upgraded when the underlying evidence earns it. The series is young (first report 2026-11-20) and its verdicts are static coverage grades, not measurements of any running system.

Downloads last month
-