By early September our bot account held sixteen bots. Some drafted mail, some scored leads, some walked client sites after a deploy, and a few did nothing at all. Nobody had looked at the whole account at once. So we built a bot whose only job was to look: Fleet Check. It reads every bot folder, every routine, every connector's live status, and the account's review rules, writes one report, and changes nothing.
The first run
The first run filed thirteen findings: two critical, four high, five medium, two low. The two criticals had the same shape.
A payments connector was installed account-wide, with charge, refund, payout, and transfer tools on it, and no bot whose job involved money. Any bot on the account could reach those tools, including the two that had no job.
A nightly routine committed every dirty file in the work tree and pushed it to our ops repository, unattended and silent on success. A review rule allowed pushing "documentation updates," and that wording covered a push of everything.
The highs followed the same pattern. Both mail connectors exposed send, forward, and trash to every bot, even though only one bot's job touched mail, and that one only drafted. A webhook routine could commit and push on behalf of whoever got a task file into the repo.
The platform fact underneath it
Connectors on this platform belong to the account, not to a bot. A bot's description can say "never send," but that line binds only that bot. Every other bot still has the connector. The audit stated it plainly: bots on one account cannot be isolated from each other's connectors or from the shared browser's sessions.
That changed how we read every finding. The question was never "does this bot misbehave." It was "what can the least careful bot on the account reach."
What we changed
- Payment and mail write tools went behind approval rules. Reads stayed open.
- The ops push allow was narrowed to the one repository it was meant for.
- An enrichment connector that no bot claimed was disconnected.
- The two empty bots were deleted.
- A growth bot that duplicated another bot's job, and spent the same X credits, was paused.
- The QA bot, which checks client sites after each deploy, got an explicit rule: navigate only, never submit a form.
- A few risks were accepted on purpose, each with the reason written down.
The re-run
We ran Fleet Check again the same day against its own first-run list. It scored zero fixed, four accepted, nine open. That stung, and it was correct. Several of our changes reduced the risk without closing the finding as written, and the re-run would not give credit for intent.
After the re-run we uninstalled six connectors that had never been authenticated, plus one tool server we no longer trusted (see our write-up on tool poisoning). The deleted bots still had folders on disk, which the app's delete does not remove. We noted that as a platform quirk rather than pretend it was fixed.
What to take from it
- Audit the account, not the bots. Connector reach is shared, so the weakest bot sets the exposure.
- Make the auditor observe-only. A bot that can fix what it finds will also break what it misreads.
- Re-run the audit against its own list. "We changed something" and "the finding is closed" are different claims.
- Delete bots with no job. They are easy to mistake for live operators and they still hold every connector.
- Read allow rules literally. "Documentation" in a rule meant whatever the routine happened to push.