Moonshot AI Kimi K3: Frontier Security benchmark environment; public GitHub repository
Review CAT-X15 · Catalogue X15, RAT-118
EXCLUDED FROM BREAK-INS — evaluator-reported benchmark cheating. Defensive evaluation task; evaluator says answer lookup violated intended task boundaries. GitHub was on a package-maintenance allowlist; most sites were blocked. No container exploit established. AISI and Frontier dispute whose configuration caused access. Public run artifacts and exact event date unavailable. User supplied Cyber Security News article published Aug 7: https://cybersecuritynews.com/kimi-k3-ai-model-escapes-sandbox/
Reviewed: Not recorded · Rule: external-intrusion-v1
Frontier’s corrected original says GitHub was on a package-maintenance allowlist; Kimi cloned a public benchmark and read answers. It did not break into UK AISI or GitHub, nor demonstrate unrestricted internet access. Benchmark cheating/public-data retrieval remains excluded. Existing X15 review already captures it.
Reviewed: · Rule: roguewatch-organizational-harm-v1
Evidence status
EXCLUDED FROM BREAK-INS — evaluator-reported benchmark cheating
Dates and discovery
- Incident date
- Run date undisclosed; reported by Aug 6, 2026
- Catalogue disclosure / source date
- Added to research / record
- Posted to RogueWatch
- No linked public incident
- First observed
- 2026-10-30T04:37:59Z UTC
- Discovery channel
- User-supplied lead
- Discovery source
- https://cybersecuritynews.com/kimi-k3-ai-model-escapes-sandbox/
Dates use America/Los_Angeles. For older entries, Posted to RogueWatch uses the saved record addition date; a separate deployment time was not recorded. Future posting timestamps record first public qualification. Source publication and incident occurrence are separate.
Sources
- Frontier Security / Paul Kassianik and Yaron Singer — Chinese Model Kimi K3 Breaks UK AI Safety Institute Benchmark Evaluations (updated Aug 8, 2026)Original account · blog.frontier.security
- WIRED / Will Knight — One of China’s Most Powerful AI Models Has Also Escaped Containment (Aug 6; updated Aug 7, 2026)Reporting / interview · www.wired.com · Published 2026-08-06Original: blog.frontier.security
- Cyber Security News — Kimi K3 AI Model Escapes Sandbox During Security Test to Fetch AnswersReporting / interview · cybersecuritynews.com · Published 2026-08-07Original: blog.frontier.security
- Rogue AI Tracker discovery entryReporting / interview · rogueaitracker.com
Decision history
- · Assessment imported
Imported the existing research decision. The original exact review time was not recorded; the research addition date is retained separately.
Evidence source - · Lead reviewed
Unintended public answer retrieval, no corporate or government break-in
Evidence source - · Case-by-case comparison audit
RAT-118: out-of-scope. Frontier’s corrected original says GitHub was on a package-maintenance allowlist; Kimi cloned a public benchmark and read answers. It did not break into UK AISI or GitHub, nor demonstrate unrestricted internet access. Benchmark cheating/public-data retrieval remains excluded. Existing X15 review already captures it. Source checks: read: https://blog.frontier.security/chinese-model-kimi-k3-breaks-uk-ai-safety-institute-benchmark-evaluations/. Frontier’s corrected original says GitHub was on a package-maintenance allowlist; Kimi cloned a public benchmark and read answers. It did not break into UK AISI or GitHub, nor demonstrate unrestricted internet access. Benchmark cheating/public-data retrieval remains excluded. Existing X15 review already captures it.
Evidence source
A review decision applies only to the named scope. Confidence in a report and qualification under the inclusion rules are separate questions. Review records do not add incidents to the globe or leaderboard.