Autonomous Security ResearchAgents that reason
We train models that understand security, build autonomous adversaries that test real systems, and engineer the controls that make AI agents safe to deploy.
Three research tracks.
AI 4 Security & Security 4 AI
Security Models
Post-training frontier models to reason like security engineers. We build the datasets, evaluations and benchmarks that measure whether a model can actually find, explain and fix what matters.
- Post-training
- Datasets
- Evals
- Reasoning
- Benchmarks
Agentic Red Team
Autonomous adversaries that plan, adapt and chain real exploits end to end. Every finding is validated and reproduced, so defenders get evidence instead of noise.
- Autonomous adversaries
- Attack agents
- Exploitation
- Validation
Agent Security
AI agents are a new class of privileged actor. We research how to give them identity, scope their tools, contain prompt injection and keep them isolated and governed in production.
- Identity / IAM
- Tool security
- Prompt injection
- Agent isolation
- Governance
Attack informs defense.
Defense informs training.
1detecteng/models2├── post-training/ SFT + RL on security tasks3├── datasets/ attack & defense traces4├── evals/ offensive + defensive suites5├── reasoning/ multi-step exploit reasoning6└── benchmarks/ reproducible scoring
How we do the work.
Adversarial by default
We assume capable, autonomous attackers exist, and build them ourselves so defenses are tested against the real thing.
Measured, not claimed
Every capability ships with an eval. If we can't benchmark it reproducibly, we don't say it works.
Validated findings only
Autonomous doesn't mean unverified. Exploits are reproduced before they are reported.
Responsible by design
Offensive research runs in scoped, isolated environments under coordinated disclosure.
Research with us.
Break things safely.
We collaborate with security teams, model developers and agent builders on evaluations, red-team engagements and joint research.
Evals · Red-team engagements · Research partnerships