Irregular is an AI security laboratory that evaluates advanced AI models for cyber-offensive capability and builds defences before those models are released. Founded in November 2023 by Dan Lahav and Omer Nevo, the firm operates at the frontier of AI safety, working directly with leading AI research labs including Anthropic, OpenAI and Google DeepMind and gaining early access to cutting-edge models. It is headquartered in Israel with a presence in the San Francisco Bay Area.
Its work centres on several products. The SOLVE framework is used to vet risks in AI models and has been adopted by the UK government and by Anthropic, including in assessing risks in Claude 4. The firm also offers proprietary cybersecurity evaluations that go beyond standard benchmarks by placing models in realistic attack scenarios; these have been cited in system cards for GPT-4, o3, o4 mini and 5. A next-generation Scenario Evaluation Suite simulates complex exploitation chains and mirrors real-world operations.
Irregular has raised $80 million and was valued at $450 million in its most recent financing. Its technical work spans AI security, offensive cyber evaluation, AI red teaming, attack simulation and risk assessment of AI models, serving frontier AI research, cybersecurity and government AI risk-assessment needs.
