Irregular
The frontier-AI security lab whose evaluations appear in model system cards
Overview
Irregular, founded in 2023 as Pattern Labs, is a security lab that red-teams frontier AI models in adversarial simulation environments, measuring both what a model could do offensively and how well it resists misuse. Its evaluations are cited in OpenAI system cards, and its SOLVE framework has been used to assess cyber risk in Anthropic models.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 1 Aug 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- Adversarial and red-team environments
- Offensive capability measurement
- Resilience testing
- Findings suitable for system-card citation
Use cases
What you can produce with Irregular
- Run a frontier model through adversarial simulation environments before it ships
- Measure offensive cyber capability — infiltration, evasion, autonomous attack behaviour — under controlled conditions
- Assess how well a model resists being driven toward misuse, not only what it can be made to do
- Apply the SOLVE framework to score cyber risk in a way labs and governments have cited
- Produce evaluation evidence that can be referenced in a published model system card
- Work alongside an internal safety team as an external party rather than replacing it
ASEAN Perspective
Irregular in Southeast Asia
ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).
A genuinely specialised outfit doing work that very few organisations can do at all, and the proof is that its findings show up in the published safety documentation of the largest labs rather than in its own marketing. If you need a frontier model's offensive cyber capability measured before it ships, the list of credible options is short and this is on it.
Two things to hold in view. The structural tension of being paid by the lab whose model you are certifying is real, openly discussed, and not unique to this firm. And evaluation environments are production infrastructure — July 2026 demonstrated what happens when one is not treated that way.
What people say
Irregular occupies a position almost nobody else does: an external party whose results turn up inside the safety documentation of the labs it tests. Its evaluations have been cited in OpenAI system cards across several model generations, and its SOLVE framework has been used to assess cyber risk in Anthropic models, with UK government involvement reported alongside. For a company founded in 2023 under the name Pattern Labs, that is unusually deep integration into how frontier models get signed off.
The work is adversarial simulation rather than checklist auditing. Models are placed in environments that probe for antivirus evasion, system infiltration and autonomous offensive behaviour, and the measurement runs in both directions — what the model could accomplish if pointed at a target, and how well it holds up when someone tries to drive it. That two-sided framing is the substantive difference from a red-team exercise that only asks whether a jailbreak exists.
Commercially it raised $80 million led by Sequoia Capital and Redpoint Ventures, with angels including the chief executives of Wiz and Eon, and reporting at the time placed it at millions in annual revenue. Google DeepMind researchers have cited it in work on evaluating emerging cyberattack capabilities.
The caveats are structural rather than about quality. This is a small field with few independent evaluators, and a lab whose findings appear in the system cards of the companies paying for the evaluation sits in an inherent tension — one the AI safety community discusses openly and which is not specific to this firm. Engagements are enterprise-scale and quote-only, so there is nothing here for an individual developer or a smaller team. And as the Anthropic evaluation incident of July 2026 showed, the environments these assessments run in are themselves infrastructure that has to be secured; that episode is a reminder that evaluation is an operational activity with its own failure modes.
Summary of public user & expert reviews, compiled by RECATOOLS.
About this listing
This entry was compiled from publicly available data including Irregular's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Irregular unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Irregular directly →
Spotted something out of date? Suggest an update →
More in Security & Safety
