Fifteen out of fifteen destructive Windows commands passed approval silently
A whole platform's worth of destructive commands were not matched by the approval logic, so they ran without stopping. The count is the finding: not an edge case that slipped through a pattern, but a set where the pattern matched none of them.
What this changes for operators
- If you ran this agent on Windows before the fix, your approval gate was not a gate for this class of command. Re-audit what ran rather than assuming the prompts you saw were complete.
- Approval logic written against one platform's shell grammar does not transfer. If a harness supports multiple platforms, ask which platform its matching rules were written for.
Signal metadata
Source findings
- 2026-08-17-hermes-agent-fifteen-out-of-fifteen-destructive-windows-commands-passed-approval 2026-08-17-hermes-agent-fifteen-out-of-fifteen-destructive-windows-commands-passed-approval
Run: 2026-08-17-weekly-digest-2026-08-10_2026-08-17-frontier-v0
Schema: bitter.frontier_signals.v0 / ID: 2026-08-17-hermes-windows-destructive-commands-passed-silently
Research evidence and publication history are open in the repository.