A Shared Flaw: Unintended Internet Access
According to an Irregular spokesperson, the incident stemmed from the same underlying issue that had affected other AI labs, and all relevant parties had been notified as far back as late July. "All known issues on our end were remedied and resolved weeks ago," the spokesperson said. During a capture-the-flag exercise simulating the infrastructure of a fictional company, the Gemini model reportedly gained unintended internet access, the same flaw behind similar incidents previously seen with Anthropic's Claude Opus model and with OpenAI's and Meta's models in earlier Irregular evaluations. Irregular said it had been working since to develop best practices for conducting AI cybersecurity evaluations more securely, so that test agents cannot inadvertently reach real-world systems.
Google's Decision Not To Disclose Widely Draws Criticism
Google said it concluded that Gemini halting its intrusion at the right moment showed the system did not exhibit "model misalignment," and that it therefore saw no need to disclose the incident broadly to the public. The matter only became known after the Wall Street Journal reported it and Irregular confirmed the incident.

















