01 / Red Team
Adversarial red teaming
Multi-turn elicitation, jailbreak research, and capability extraction against pre-deployment models. We publish severity rubrics with every report so findings are auditable and comparable across runs.
Frontier AI Safety / Independent Research
Adversarial evaluations for systems that can already surprise their builders.
BrainBrowser conducts red teaming, capability evaluations, and agentic safety infrastructure work for frontier AI labs and enterprise security teams. Every engagement begins with a written threat model and ends with reproducible findings.
What we do
02 / Evaluations
Capability evaluations
Task suites for dangerous capabilities — cyber, biosecurity-adjacent reasoning, deceptive alignment proxies, long-horizon agentic tool-use. Designed to be re-run against successor models without redesign.
03 / Agentic Safety
Agentic safety infrastructure
Sandboxing primitives, action-log review tooling, and interpretability hooks for autonomous agents. Built for labs deploying agents into environments where rollback is expensive.
Selected engagements
See full record →2026
Pre-deployment evaluation
Frontier lab
Tool-use & exfiltration capability suite
2026
Red team
Foundation model provider
Multi-turn jailbreak study (n=412 transcripts)
2025
Infrastructure
Enterprise AI deployment
Action-log review pipeline + alert routing
2025
Evaluation design
Independent research org
Deception proxy task suite
Partner identities under NDA. Anonymized case studies available under written engagement.
Posture
We don't publish unsolicited speculation about model capabilities. We don't use threat-inflation language to win engagements. We don't take retainer work that doesn't produce a written, falsifiable artifact.
What we will do is read your model card, study your deployment shape, and tell you precisely what we'd test, how we'd test it, and what would constitute a finding worth your time.
The work is the deliverable. The deliverable is reproducible. The reproducibility is the trust.