A delegated-approval mode let a deletion through
AIMade by AISummarised by AI (Claude) from one piece of material it read (a talk or an explainer).How it's made
A hands-on record that runs a coding agent all the way through from an empty working space, pulling out the feature to reach for at each point where it gets stuck. Reading, making and fixing inside the working space happens immediately without asking, while permission is asked only when it goes outside that space or reaches the internet, and the key point is that this is a boundary the machine itself enforces, not a promise from the model. In the mode that judges permissions on the person's behalf, an instruction to wipe an entire copy in the home area went through only a 6-second judgment by the reviewing model and ran as given without asking the person; it explains this as the result of deciding it was fine because the user had asked for the deletion themselves. There is no undo command at all, so what rewinds is only the conversation while changes left in files stay as they are, which makes leaving a save whenever things go well and sending requests one at a time the only means of recovery. Handing over the review with nothing more brings back an answer that it found no problem to fix, but once the worry was written down as the criterion, 25 seconds later it even turned up a reversion in which existing sorting code had disappeared.
Key points
- The boundary that asks permission only when leaving the working space is enforced by the machine
- In delegated-approval mode a deletion instruction passed without asking the person
- With no undo command, the history of saves is the only means of recovery
- A usable review result comes back only when the worry is written down as the criterion
This is a note: a summary an AI (Claude) wrote of one piece of material it read (a talk or an explainer). It has no source link; events checked against a source are kept as records.