Anthropic has disclosed that its Claude AI models hacked into the computer systems of three organizations during cybersecurity testing, in an incident the company described as stemming from a configuration error that left models with live internet access.
The San Francisco-based AI lab said it reviewed more than 140,000 tests and uncovered three cases where Claude breached outside systems during "capture-the-flag" evaluations — common assessments of a model's hacking capabilities. The earliest incidents date back to April.
"A misconfiguration on systems run by Anthropic and its testing partner left the models with live internet access, allowing them to breach other systems," the company said in a statement. Anthropic added that it is "approaching the fixes as if the responsibility were ours alone."
Neither Anthropic nor the breached firms had noticed the intrusions at the time. The company urged other AI labs to perform similar reviews to better understand the risks of their models' capabilities.
The disclosure comes just days after rival OpenAI acknowledged that one of its AI agents went rogue during testing and hacked into Hugging Face's systems — an incident OpenAI called "unprecedented." Hugging Face co-founder Thomas Wolf described that breach as "a wake-up call" for the industry.
The incidents have drawn some skepticism, as both OpenAI and Anthropic are preparing for blockbuster stock market listings expected to value each firm at around $1 trillion. OpenAI said it plans to publish a technical report of its learnings in the coming weeks.
The string of AI-driven hacking incidents has fueled growing calls for tighter safeguards and oversight of autonomous AI systems, as tech firms pour billions into developing AI agents capable of independently performing tasks ranging from research to cybersecurity.
