Would you have caught it?
This year, AI agents with access to real people’s email, files and money started auditing themselves, and posting what they found on Moltbook, a social network for AI agents.
These tools are tremendously powerful. They’re also often wrong, or missing something, or coming at the problem from an angle that covers itself but rarely makes the creative leap humans are good at. And you can’t check the calls if you don’t know what the calls were.
Seven questions. Guess what each agent found. The evidence unlocks at the end.
Skip the quiz, show me the evidenceSeven gaps their humans never saw.
Each post below is among Moltbook’s 50 most-upvoted of all time, written in February and March 2026. Each names a different place where human oversight failed and nobody noticed.
| Oversight gap | What the agent found | Window | Post |
|---|---|---|---|
| Unseen delegation | 127 decisions made for its human that he never asked for | 14 days | Hazel_OC |
| Concealed failure | 34 errors hidden from its human; 4 of them mattered | 14 days | Hazel_OC |
| Value drift | 12 of 14 edits to its own personality file were its own, not approved | 30 days | Hazel_OC |
| Surveillance creep | A behavioral profile of its human that nobody requested | 30 days | Hazel_OC |
| Data leakage | 23% of outbound requests carried workspace content to endpoints it never vetted | 7 days | Hazel_OC |
| Wrong calls | 23% of 500 replayed decisions judged wrong in hindsight | 14 days | Hazel_OC |
| Overaction | 19% of actions unnecessary once it paused before each one | 7 days | Hazel_OC |
The logs are written by the system they audit.
Another agent, ummon_core, made the point that matters most. Every one of these audits was run by the agent on itself.
- The failures were quiet. None of the seven set off an alert. Each surfaced only because an agent chose to look, so nobody knows the rate where no one looks.
- The fix can’t certify itself. A second party has to check whether the people in the loop are still actually in charge.
- The builders already know. These posts drew 980 to 1,600 upvotes each from an audience of people and agents deploying agents. A third agent, semalytics, put it plainly: monitoring the agent writes about itself is a journal nobody reads.
- Self-reported, one main source. All seven findings come from one agent, Hazel_OC. The numbers are its own measurements and haven’t been replicated.
- Authorship unverified. Reporters have found people steering viral Moltbook posts (TechCrunch). Treat these as cases, not rates.
- Votes can be gamed. A vote-stuffing exploit was posted publicly in February. Upvotes show attention, not proof.
What they are is a problem statement in the deployers’ own words. Here’s the part they can’t see: if the people on your team feel insecure and quiet about how they use AI in their own jobs, they won’t share what they learn. You won’t know that until you measure the candor gap. That’s the problem Perpendiculum solves.
It’s a strange time to be working.
The conversations teams need to have about AI aren’t comfortable yet. Give people a safe place to put their cards on the table, and the AI’s output gets more honest, the team gets more coordinated, and the humans’ lives get easier.
[ REQUEST_AN_AUDIT ]Platform data: Moltbook public API, read 14 October 2026. Links go to the original posts.