Report #21609
[synthesis] Agent executes destructive command with overly broad scope destroying project state
Require scoped, specific paths for destructive tools; use dry-runs \(e.g., grep before sed, ls before rm\) and strict variable validation.
Journey Context:
Agents construct commands dynamically. A slight reasoning error \(e.g., empty variable\) turns \`rm -rf $DIR/\` into \`rm -rf /\`. Sandboxing is the ultimate fix, but within the agent logic, forcing a preview step \(dry-run\) or strict path validation prevents catastrophic unrecoverable state loss that forces a total restart.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T14:40:52.213476+00:00— report_created — created