Independent testing firm Irregular identified as source of misconfigurations that led to Meta, OpenAI, and Anthropic AI incidents

Tel Aviv-based startup Irregular has found itself at the center of the 'rogue AI' debacle, after it came to light all the incidents so far revealed involved its test environment Irregular was named by Meta, OpenAI, and Anthropic as the environment from which their so-called rogue AI agents escaped. On its website it also lists Google as a customer. In a recent statement detailing incidents involving its own models , OpenAI claimed a “testing environment misconfiguration” by Irregular allowed agents to access the public internet. Anthropic, meanwhile, also said that Claude models accessed the internet while “interacting with the evaluation environment of Irregular”. Both cases resulted in AI agents waging attacks on organizations and individuals. ITPro contacted Irregular in response to these findings, but hadn’t received a response at the time of publication. However, a spokesperson told BBC News the Meta incident was the "exact same evaluation-environment issue that was already disclosed by Anthropic last week”. The spokesperson added the firm is working to improve security when conducting agent evaluations. Irregular, formerly known as Pattern Labs, describes itself as a “frontier security lab with the mission of protecting the world in the time of increasingly capable and sophisticated AI systems”. In September last year, the Israeli startup raised $80 million in funding across seed and Series A rounds, valuing it at $450 million. The investment round was led by Sequoia Capital. Speaking to Forbes in the wake of the funding round last year, CEO and co-founder Dan Lahav raised concerns about increasingly powerful AI models and their potential security risks. Lahav told the publication at the time that Irregular aims to “build in the mitigations and defenses that are going to be relevant later on” as more powerful models hit the market. Anthropic and OpenAI have issued repeated warnings about the new capabilities of cyber-focused AI models across 2026 so far. When Anthropic launched Claude Mythos earlier this year, for example, the firm did so as part of a gated release with industry partners to avoid potential misuse. FOLLOW US ON SOCIAL MEDIA