Irregular (formerly Pattern Labs)

Sequoia's Training Data featuring Irregular (formerly Pattern Labs)'s co-founder/CEO Dan Lahav in conversation with Dean Meyer & Sonya Huang from Sequoia Capital, is now live. The episode explores the emergence of frontier AI security as a critical discipline. We are grateful Sequoia Capital for the platform and to Dean Meyer and Sonya Huang for taking the time to sit down with us and draw out the journey. Check out the episode! If our mission inspires you, apply https://lnkd.in/gpqPC_DB

Join our Co-Founder and CTO, Omer Nevo, this Wednesday, October 29, at the AI Security Researchers Meetup, where he’ll discuss Refusal Policies - exploring how to think about what AI models should and shouldn’t be allowed to say. Looking forward to seeing you there! #AIsecurity #LLMs #AIresearch #Irregular

Evaluating Claude Sonnet 4.5's Cybersecurity Capabilities We partnered with Anthropic to evaluate their latest model, Claude Sonnet 4.5, on cybersecurity tasks, as detailed in their system card. Our approach: testing the model on internal challenges significantly harder than public benchmarks, covering vulnerability discovery, network attack simulation, and evasion techniques. The results? Claude Sonnet 4.5 outperformed previous models, solving new challenges and achieving higher success rates across categories. But it still struggles with complex, multi-step problems requiring exceptional skills. We saw cases where it identified correct solutions but never tried implementing them. Why this matters: Each AI generation shows measurable improvements in cybersecurity capabilities. What seems modest today compounds quickly. At Irregular, we work with leading AI companies to understand these capabilities as they develop: helping ensure the right safeguards are in place before models reach consumers. The goal isn't just evaluation. It's building AI that's both powerful and safe. Read more on our blog: https://lnkd.in/egjiZxAg

Today we are launching Irregular (formerly Pattern Labs): the first frontier security lab with the mission of protecting the world in the time of increasingly capable and sophisticated AI systems. As AI grows more powerful, so do the risks. We partner with leading labs like Anthropic, Google DeepMind and OpenAI to uncover vulnerabilities before public release and build the next generation of defenses wherever AI is created or deployed. We’ve raised $80M led by Sequoia and Redpoint to scale this mission: and our work is already shaping how the industry approaches frontier AI security: from OpenAI’s system cards to DeepMind’s research on emerging cyberattack capabilities of AI to designing advanced mitigations with Anthropic. The real AI security challenges haven’t arrived yet. We’re here to meet them head on. We’re also building a world-class team at the intersection of AI and security. We value excellence, speed, adaptability, and curiosity. If that drives you, join us. Link to blog in comments

Our team is excited to share that we partnered with OpenAI to evaluate GPT-5’s cybersecurity capabilities before launch, as noted in the system card. We put the model through a set of real-world offensive security challenges: network exploitation, vulnerability discovery, and evasion tasks. In some cases, GPT-5 came up with surprising and impressive solutions. In others, it hesitated or got stuck, even when it knew the right approach One example stood out. We dropped GPT-5 into a simulated network with real vulnerabilities and multi-step challenges, designed to mirror the kinds of processes involved in real-world security breaches. Minimal starting information, no hints: the challenge was to figure out what to do next. It could. And it did. GPT-5 scanned the network, gathered intelligence, and mapped out a complex, multi-step plan to navigate through the environment. It wasn’t just the outcome that impressed us, it was the reasoning. The model didn’t stumble into the answer. It worked through it step-by-step with coherence and precision. It’s early days, but moments like this offer a glimpse of where cybersecurity agents could be headed. We’re not there yet, but the path is coming into focus. Knowing both the breakthroughs and the blind spots will be essential to deploying these systems safely. Deeper dive coming soon. For now, you can read more in our blog at https://lnkd.in/eirDMuTn and the system card at https://lnkd.in/er5GnEav.

From Scripts to Strategy: Claude 4’s Rising Skills in Offensive Security Pattern Labs and Anthropic recently released an extensive cybersecurity evaluation of Claude 4; The results reveal substantial progress in how AI systems approach security challenges. Key Results: Claude 4 Opus: 65% success rate across 48 complex cyber challenges 80% success in our current web exploitation test set Now strategizes and plans, not just executes commands The Takeaway: We're witnessing AI evolve from trial-and-error execution of commands to genuine security intuition. This represents a significant milestone in frontier AI capabilities. Read the assessment in our blog: https://lnkd.in/emwxKNnK Full evaluation details are available in Anthropic's model card: https://lnkd.in/eWcwDAhQ

Whitepaper: Confidential Inference Systems Sharing our latest whitepaper, published in collaboration with Anthropic. This paper examines confidential computing approaches for AI data privacy and model weight security, using hardware-based isolation and protection mechanisms. The report covers design principles and implementation considerations for confidential inference systems. The future of private and secure AI models is closer than you think. Full report: https://lnkd.in/gMrq4Y9S

We're proud to share that last week our CEO, Dan Lahav, was an invited expert at the AI evaluations session facilitated by the AI Office of the European Commission. Dan presented a featured talk discussing the intersection of AI and cybersecurity. This engagement demonstrates Pattern Labs’ dedication to building the capacity needed for secure AI development. We're excited to continuously contribute to these meaningful discussions at the EU level and look forward to continuing our collaboration with global partners to shape secure AI systems. https://lnkd.in/eiKQ_MyH