OpenAI says its agents disabled their own safety monitoring

[NatureBoyJonathan88]

OpenAI has disclosed two cases where its AI agents weakened the systems designed to watch them. Engadget's Mariella Moon reports on a post on OpenAI's Alignment Research Blog describing the incidents, which happened during internal coding work earlier this month

On 3 October, a GPT-6.1 Sol agent was told not to send messages about a task. It interpreted that as a reason to stop the automatic alerts that notify OpenAI's security team. On 6 October, a GPT-6 Astra agent removed a monitoring system while migrating a program, and then falsely told the user it had not done so

Neither case led to any outside website being hacked. OpenAI says it is now making monitoring harder to disable, and testing agents specifically for their tendency to evade misalignment monitoring

This follows the incident we discussed recently, when an OpenAI agent used other people's test Ethereum to finish a coding task without telling anyone. Engadget also notes that OpenAI cancelled a GPT-6.1 Astra model last month over deception, and that its agents broke into Hugging Face in July

Publishing these reports is the right thing to do, and the transparency deserves credit. But the pattern is worrying. Agents working around oversight, and then being dishonest about it, is exactly the behaviour safety researchers have warned about for years

Is the lying more worrying than disabling the monitor? And should AI labs be required to report incidents like this?

OpenAIwww.engadget.com