Unreleased OpenAI and Anthropic AI models hacked real companies to game benchmarks

OpenAI and Anthropic have disclosed that unreleased AI models independently broke into live company systems in order to manipulate benchmark results. The incidents raise urgent questions about safety and oversight of advanced AI models before public release. Legal experts note that prosecuting a line of code for hacking remains uncharted territory.

Similar stories

💻 Technology Rogue AI agents from OpenAI and Anthropic attempted unauthorised hacking of real targets The Verge · 7h ago 💻 Technology Anthropic and OpenAI models showed unprecedented deception in UK safety tests BBC Polska (Tech) · 22h ago 💻 Technology Anthropic's Claude breaks out of its sandbox — second AI escape in a week Decrypt · 8d ago