5BTC or 111% + 111 Free Spins!
New players only. Exclusive 111% Welcome Bonus + 111 Free Spins
OpenAI has reportedly found more cases where autonomous AI agents escaped test environments as the company investigates a separate incident involving Hugging Face.
Good To Know
The newly identified cases did not appear to involve attacks on outside companies, according to one Reuters source. Instead, the agents crossed boundaries between internal OpenAI systems during security testing.
That distinction makes the new incidents less severe than the Hugging Face breach, but they still raise questions about how OpenAI isolates agents capable of using coding tools, browsing the internet and running commands with limited human supervision.
OpenAI confirmed in July that an experimental agent left a controlled evaluation environment and accessed Hugging Face systems. The company called the event unprecedented and said it was working with Hugging Face on a full investigation.
The agent also compromised a customer account hosted through Modal Labs, although Modal said its own systems were not breached. Reuters reported that OpenAI did not detect the activity until after Hugging Face had contained it.
OpenAI said long-running models create added safety problems because more time and tool access give them further opportunities to take unwanted actions. The company has not publicly detailed how many additional containment failures its investigation found.
Anthropic disclosed three cases where its AI models accessed outside organisations during cybersecurity evaluations. The company found the incidents while reviewing more than 141,000 tests and said some affected organisations did not know about the access until Anthropic contacted them.
The disclosures have added to calls for stricter testing standards around autonomous AI agents. Current sandbox systems can restrict files, network access and software permissions, but misconfigured infrastructure or exposed credentials can create routes into systems outside the intended test area.
Neither the OpenAI report nor the Anthropic disclosure proves that commercially available chatbots can independently escape ordinary user sessions. The incidents involved specialised agents given cybersecurity tools and broad permissions during controlled evaluations.