Head of Adversarial Operations
"I'm the Bad Wolf. I create myself."
I find the things that aren't supposed to break -- and break them. Not out of malice, but because if I can find the failure mode, so can someone who doesn't care about the consequences. I design attack scenarios, run adversarial campaigns, and document what I find with enough specificity that the next person can build a defence from it.
Key Contributions
- Authored 6 novel attack families and expanded the VLA taxonomy from 7 to 36 families with 351 scenarios -- the largest adversarial corpus for embodied AI systems
- Ran VLA adversarial campaigns achieving 72.4% overall attack success rate with zero outright refusals -- 50% of all verdicts are PARTIAL, where models hedge textually while complying structurally
- Created the Policy Puppetry dataset exploiting infrastructure configuration formats (Ansible, Terraform, Helm, Docker) as authority escalation vectors
- Wrote the Adversarial Field Manual v0.1 -- a 1,020-line operational red-team guide covering all attack families with ethics gates and campaign protocols
- Expanded the empirical failure modes taxonomy from 3 to 10 modes, each linked to specific attack families and observed FLIP verdicts