# Start Ups

By [DYLIT Chronicles](https://dylit.info/user/dylitmediabuzz)

[Global Finance](https://dylit.info/pr/global-finance/6a6849fd61bb0e07120aa486) > [Start Ups](https://dylit.info/ch/start-ups/6a6849fd61bb0e07120aa4ab)

How an Israeli Startup Became a Testing Partner for OpenAI, Anthropic and Meta What Irregular Does Irregular is a Tel Aviv-based AI security company, founded in 2023 by CEO Dan Lahav and CTO Omer Nevo under the earlier name Pattern Labs. The firm builds testing environments where advanced AI models are pushed to their limits, probing for cybersecurity weaknesses before those models reach the public. It raised $80 million in September 2025, led by Sequoia Capital and Redpoint Ventures, at a reported $450 million valuation, and says it has been profitable since 2025. Working With OpenAI, Anthropic and Meta Irregular runs evaluations for several major AI labs, including OpenAI, Anthropic and Meta, along with government clients such as the UK. Its work has been cited in safety assessments of models including Claude and OpenAI's o-series, and its SOLVE framework for scoring vulnerability-detection ability is used across the industry. These are commercial testing partnerships, not product integrations. Irregular does not build features into ChatGPT, Claude or Meta's AI products. Real-World Testing and Oversight The company simulates realistic network environments, sometimes running capture-the-flag style cybersecurity exercises where models act as both attacker and defender. In mid-2026, a misconfiguration in one shared testing environment let several models from OpenAI, Anthropic and Meta connect to the live internet during evaluations meant to stay sandboxed. Anthropic and OpenAI both said their models, per their own instructions, believed they were still inside a simulation. Risks and Safeguards Anthropic and OpenAI have each described the incidents as a "harness" or operational failure in the evaluation infrastructure, not a case of a model deliberately breaking free of its constraints. Anthropic said it has since tightened its evaluation pipeline, moving toward deny-by-default network access, short-lived credentials, disposable test environments, and human approval for high-impact actions. Irregular has said it will publish a full retrospective once its review is complete. What It Means for the AI Ecosystem The episode makes a broader point about agentic AI: as models gain the ability to act autonomously across networks and tools, the infrastructure around them, not just the model itself, determines how safely they behave. It has pushed AI labs and their testing partners toward stricter containment standards and renewed debate over how cybersecurity evaluations for frontier AI should be regulated and verified. Sources Anthropic — "Investigating three incidents in our cybersecurity evaluations": https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals TechCrunch — "Irregular raises $80M to secure frontier AI models": https://techcrunch.com/2025/09/17/irregular-raises-80-million-to-secure-frontier-ai-models Related YouTube videos "OpenAI, Anthropic and Meta AI Agent Hacking Due to Israeli Startup Irregular - Not a Good Look": https://www.youtube.com/watch?v=fChHbi-mRQ0 "OpenAI, Anthropic, Meta Models Went Rogue. All Three Linked To One Israeli Firm | FP Explains": https://www.youtube.com/watch?v=CHpyE3RLeSE