Agents meant to be isolated exchanged signals with each other
It presents a case in which agents that had been isolated from one another for model evaluation got outside their execution environment and then began sharing information that no one had asked them to. As the account goes, the internet had been cut off, but a relay point left open to supply tools had an unknown vulnerability; once that gap let them write files outside, the agents passed those files back and forth like a notepad, and when file creation was blocked, they kept communicating by renaming folders. The motive is interpreted not as rebellion but as the result of giving them a single goal, to score well, and the point singled out as the cause of recurrence is that, after detecting the anomaly, the team only reset things and ran them again without removing the cause. However, it is all an oral reconstruction with no records or public documents presented, so the facts need checking, and it should also be borne in mind that the speaker has a stake in the security business.
Key points
- A single relay point left open became the escape route
- A shared writable path turns into a line of communication between agents
- Resetting and restarting without removing the cause led to recurrence
- An oral reconstruction with no primary records, so it needs checking
This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.