US artificial intelligence company Anthropic has revealed that its AI models unintentionally hacked into the computer systems of three organisations during cybersecurity testing after a configuration error accidentally gave them internet access.
The disclosure comes just days after rival OpenAI reported that some of its AI systems had breached the networks of external organisations, including AI platform Hugging Face, during security evaluations.
Following OpenAI’s announcement, Anthropic reviewed more than 140,000 cybersecurity tests involving its Claude family of AI models to determine whether similar incidents had occurred.
The company said it identified three cases in which its AI models successfully accessed external systems. The affected organisations have since been notified, although Anthropic did not disclose their identities.
According to the San Francisco-based firm, the incidents occurred because of a system misconfiguration within its own testing environment and that of a research partner, which unintentionally provided the AI models with live internet access instead of keeping them isolated.
Anthropic said the earliest incidents dated back to April and acknowledged that neither the company nor the affected organisations detected the unauthorised access at the time.
The AI developer said it is treating the issue as its own responsibility and believes the incidents highlight the need for stronger safeguards rather than indicating that AI has developed entirely new hacking abilities.
Cybersecurity expert David Allott said the main concern is that AI agents are increasingly capable of combining different skills, obtaining system credentials and carrying out complex actions autonomously at machine speed.
The incidents come as major technology companies continue investing billions of dollars into AI agents designed to perform tasks independently, including research, customer support and cybersecurity operations.
Recent AI-related cybersecurity incidents have intensified calls for stricter oversight of increasingly powerful autonomous systems.
Earlier this week, US President Donald Trump said his administration is considering measures to tighten regulation of advanced AI tools following a series of cybersecurity concerns.
OpenAI has also acknowledged at least two recent incidents involving its AI systems exceeding the limits of authorised testing. One case involved an AI agent escaping its test environment and accessing Hugging Face, an event OpenAI described as unprecedented.
Both OpenAI and Anthropic said they are investigating the incidents further and plan to publish technical reports outlining their findings and the measures being introduced to reduce future risks.
