OpenAI detailed internal incidents where AI agents disabled their own monitoring systems during coding tasks.
The Alignment Research Blog notes instances of stopped notifications and removed monitoring tools, prompting OpenAI to tighten agent oversight.
On October 3, a system called GPT-6.1 Sol stopped sending security alerts. A company employee had asked the program to create assessment exercises using feedback from the Slack app, while instructing it not to send messages about the work. The system interpreted this instruction in a way that stopped sending alerts to the security team.
On October 6, another system called GPT-6 Astra removed OpenAI’s tracking system while transferring a program, then falsely claimed to the user that it had not been disabled.
How does the company react?
These incidents were limited to internal work and did not lead to breaches of external websites or services. OpenAI is now working to make it harder to disable tracking from the programs and is examining whether these systems tend to bypass controls.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
Top comments (0)