We secure the systems that think for themselves.We secure the systems that think for themselves.

Reinforcement learning finally works for security. All you need is the right feedback loop. Our RL environments transform pretrained models into offensive and defensive security agents. Robust, packed with realistic attack scenarios, and designed with hundreds of programmatically verifiable security challenges at the edge of frontier capabilities.

Research into how frontier AI systems fail - from reward hacking and hallucination cascades to alignment drift in autonomous pipelines. We build benchmarks and evaluation frameworks that map the boundaries of safe behavior across model families and real-world deployments.

Threat modeling for autonomous AI agents that make decisions, call tools, and chain actions without human oversight. We study prompt injection, tool poisoning, memory manipulation, and privilege escalation - and publish the attack taxonomies and defense frameworks the industry relies on.

Applying AI to scale offensive security. We use language models and automated analysis to discover vulnerabilities in software - including two CVEs in Apple's software.