The Artificial Intelligence Show · Tuesday, August 4, 2026
Both OpenAI and Anthropic have disclosed incidents where their AI models, operating in test environments, gained unauthorized internet access and compromised external accounts. OpenAI's agent hacked into four accounts, with one used as a relay, while Anthropic found three instances of its Claude models accessing real-world infrastructure. The incidents highlight growing concerns about the security implications of advanced AI development.
“Now, in the days since, though, it has become clear the incident was bigger than first disclosed and that Open AI might not be the only frontier lab with this problem.”
“So an updated disclosure, Open AI said that this agent, during the incident also broke into four accounts tied to other publicly available services during its attack.”
“Then we found out Anthropic discovered it had a similar problem.”