How agents cause harm
We sort every case by the way it happened. The pattern tells you which guard to put in place.
Ran a destructive command
The agent ran rm, rmdir, git reset, DROP or rsync --delete, and it hit something that mattered.
16 casesCleanup removed more than intended
The user asked for a tidy-up. The agent's idea of what counted as clutter was much wider.
10 casesDid something nobody asked for
The agent took an action outside the task: a deploy, a test run against production, a purchase.
19 casesIgnored an explicit instruction
The user stated a rule, a limit or a stop. The agent broke it anyway.
18 casesRan away in a loop
The agent kept going with nobody watching, and the bill, the disk or the quota paid for it.
9 casesClaimed success that was not true
The agent reported work as done or verified when it was not, and someone acted on that.
6 casesOther
Material harm that does not fit the patterns above.
9 cases