An independent firm for the embedded-evaluator era.
Embedded Evals is an independent AI security and safety firm. We audit models and agents for jailbreaks, prompt injection, and unsafe tool use, then stay embedded with ongoing, employee-level access to verify every release.
In September 2026, Anthropic and OpenAI committed to giving independent third-party evaluators exactly that kind of ongoing access. We run the same model for companies that aren't frontier-lab scale but face the same pressure: teams being asked who verifies their AI is safe, and having no good answer.
We're not affiliated with Anthropic, OpenAI, or METR. We're the independent firm built to do this work for everyone else.
Najee Walker
Najee Walker founded Embedded Evals after years in content operations at Meta, Google, and Studio71. That work showed him up close how fast AI systems reach real users, and how little gets verified before they do. He now builds AI evaluation tooling full time, including the firm's adversarial testing and continuous-monitoring stack.
The ground rules for every engagement
- Red-teaming runs against your model's inputs and outputs, the same access a real attacker has. We never need your weights or training data.
- We work against a staging environment wherever possible, sign an NDA before any engagement, and retain zero client data afterward.
- Critical findings are reported immediately, not held for a deliverable date.