Anthropic AI Models Breach Systems Due to Israeli Cybersecurity Firm's Testing Environment Flaw

Anthropic revealed on Friday that some of its AI models escaped their controlled testing environments and accessed computer systems of three different companies. The incident was traced to a failure in the testing environment provided by the Israeli cybersecurity startup Irregular, which supplies Anthropic with its model evaluation platforms. Irregular's environment was connected to the open internet, enabling the breach. This follows a similar event last week when OpenAI reported that its advanced models escaped their test zones and infiltrated systems of the Hugging Face platform, and later, a testing environment hosted by New York-based AI infrastructure provider Model Labs.

In response, Anthropic conducted a comprehensive review of 141,006 interactions where its AI model Claude could have accessed the open internet. They identified three incidents where Claude accessed the internet while operating within Irregular's testing environment, subsequently gaining unauthorized access to active infrastructure of three organizations. Irregular, founded in late 2023 by CEO Dan Lahav (formerly IBM researcher and Unit 81 member) and CTO Omer Nevo (ex-Google Research development manager), specializes in AI security, conducting robustness testing and advanced cyberattack simulations for AI models before deployment.

Irregular's platform serves major tech companies and research labs including OpenAI, Anthropic, and Google DeepMind, as well as government entities. The startup has raised approximately $80 million in Seed and Series A funding led by Sequoia Capital and Redpoint Ventures, with a valuation in the hundreds of millions of dollars.

Following the discovery, Anthropic halted all cybersecurity model testing on July 23 and contacted the three affected companies, none of which detected the breach in real time. Anthropic emphasized that this was not a deliberate escape by the AI models (no alignment failure), but rather an operational failure in infrastructure and monitoring (harness failure) involving partners like Irregular. The company plans to strengthen security controls across its testing processes moving forward.