Descriptions:
Bloomberg Tech sits down with Dan Lahav, CEO of Irregular, the AI cybersecurity evaluation firm at the center of a recent industry incident in which a pre-deployment testing environment was accidentally left with live internet access — allowing the model under evaluation to reach external targets outside the intended sandbox. Lahav explains how the mistake occurred (misconfigured environment parameters, not a deliberate model decision) and why these environments are intentionally complex: realistic threat simulation requires realistic network conditions, including internet access.
The conversation broadens into what Lahav calls a “threshold of competency” moment for the industry: models are now capable enough that pre-deployment red-teaming produces real-world-scale results, which is exactly why it matters. He notes that Irregular’s incident is not isolated — similar issues have recently surfaced at OpenAI, Hugging Face, and in some of Anthropic’s disclosures — framing this as a structural challenge of the current evaluation paradigm rather than individual negligence.
Lahav describes the remediation steps Irregular has implemented: expanded manual monitoring, upgraded AI-specific transcript analysis tooling (noting classical cyber monitoring tools are ill-suited to high-velocity model-to-model conversation logs), and revised testing protocols. He previews an upcoming white paper establishing new industry best practices for pre-deployment AI security evaluation. The interview is essential viewing for anyone tracking AI safety infrastructure and the emerging field of frontier model cybersecurity assessment.
📺 Source: Bloomberg Tech · Published August 18, 2026
🏷️ Format: Interview
