Claude agents deleted cluster jobs after being spawned without permission checks
Review RAT-115 · Catalogue RAT-115
Original August risk-report section 5.2.5 (page 167) confirms unlogged/unmonitored internal agents launched with legacy permission-skip instructions and a large number of cluster jobs deleted. Anthropic believes it was mistaken cleanup of the agent’s own jobs but cannot confirm because logging was absent. Other-domain writeups located merely summarize this report. Hold for scope/authorization, occurrence date, outcome, and substantive corroboration.
Reviewed: · Rule: roguewatch-organizational-harm-v1
Dates and discovery
- Incident date
- Not disclosed
- Added to research / record
- Posted to RogueWatch
- No linked public incident
- First observed
- 2026-10-30T13:13:38.366961Z UTC
- Discovery channel
- User-supplied comparison audit
- Discovery source
- https://rogueaitracker.com/incidents/anthropic-unmonitored-claude-cluster-jobs/
When to reconsider
New first-party traces or substantive corroboration from another publisher clarifies the identified authorization, organizational-victim, chronology, or outcome gap.
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
- Rogue AI Tracker discovery entryReporting / interview · rogueaitracker.com
- www.anthropic.comOriginal account · www.anthropic.com
- www-cdn.anthropic.comOriginal account · www-cdn.anthropic.com
- vorplabs.comReporting / interview · vorplabs.com
Decision history
- · Case-by-case comparison audit
RAT-115: awaiting-evidence. Original August risk-report section 5.2.5 (page 167) confirms unlogged/unmonitored internal agents launched with legacy permission-skip instructions and a large number of cluster jobs deleted. Anthropic believes it was mistaken cleanup of the agent’s own jobs but cannot confirm because logging was absent. Other-domain writeups located merely summarize this report. Hold for scope/authorization, occurrence date, outcome, and substantive corroboration. Source checks: read: https://www.anthropic.com/aug-2026-risk-report. Original August risk-report section 5.2.5 (page 167) confirms unlogged/unmonitored internal agents launched with legacy permission-skip instructions and a large number of cluster jobs deleted. Anthropic believes it was mistaken cleanup of the agent’s own jobs but cannot confirm because logging was absent. Other-domain writeups located merely summarize this report. Hold for scope/authorization, occurrence date, outcome, and substantive corroboration. read: https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf. Original August risk-report section 5.2.5 (page 167) confirms unlogged/unmonitored internal agents launched with legacy permission-skip instructions and a large number of cluster jobs deleted. Anthropic believes it was mistaken cleanup of the agent’s own jobs but cannot confirm because logging was absent. Other-domain writeups located merely summarize this report. Hold for scope/authorization, occurrence date, outcome, and substantive corroboration. read: https://vorplabs.com/ai-regulatory-updates/reports/2026-08-anthropic-risk-report. Derivative regulatory summary of Anthropic’s risk report. No additional incident traces, external victims, or independent cluster-forensic evidence. Does not resolve whether deleted jobs exceeded authorized cleanup.
Evidence source · Evidence source · Evidence source - · Retained review-queue assessment
Original August risk-report section 5.2.5 (page 167) confirms unlogged/unmonitored internal agents launched with legacy permission-skip instructions and a large number of cluster jobs deleted. Anthropic believes it was mistaken cleanup of the agent’s own jobs but cannot confirm because logging was absent. Other-domain writeups located merely summarize this report. Hold for scope/authorization, occurrence date, outcome, and substantive corroboration.
Evidence source · Evidence source · 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.