Security architect for AI agents
CISSP · M.S. Artificial Intelligence, Santa Clara University · Navy veteran
The helpfulness is the hazard. A capable agent without context or bounds does not refuse. It fills the gap with a guess and acts, with real permissions.
So the control cannot live inside the model. I build the control that sits in front of every action an AI agent takes, and answers three questions: who allowed this action, how do we stop it now, and how do we prove what happened.
capability-gate decides each tool call before it runs. Deny by default, a kill
switch, and each decision on a hash-chained log before the answer goes back to the agent.
One command shows it working, and 34 checks each delete one control and confirm a
test catches it. It checks tool names and paths, not shell command text.
aegis-atoms holds small policy facts and a judge that may tighten a verdict, never loosen it. Built and tested; not measured live.
isolation-layer runs a risky task in its own Firecracker microVM under the jailer. Research build. Jailed boots on a separate KVM test host are logged in its receipts; CI runs a stand-in VM. Wiring the gate's isolate route to the real box is the next step.
Research. owasp-dual-top10-lab maps the OWASP Top 10 for LLM and agentic apps to MITRE ATLAS techniques, CVEs and AI incidents. newwave-owasp-security-lab is the earlier course lab.
- Controls ship with a test that fails when the control is removed.
- Observe before enforce: watch real traffic, then write the policy.
- The limit sits next to the claim.
- Independent review before anything merges.
Nearly seven years of Navy security operations across Bahrain, Poland and Norfolk. Built the security program for a NATO installation from construction to commissioning. Led a fifty-person security department at Naval Station Norfolk as Security Operations Manager (billet: Watch Commander).
M.S. Artificial Intelligence at Santa Clara University, in progress. B.S. Cybersecurity Technology. CISSP, ISC2 CC, CompTIA Security+ and A+.
Earlier work
cpt-reimbursement-engine. A dependency-free TypeScript engine that predicts where Medicare reimbursement fails, not just whether a code exists.
crypto-signal-confluence. A research instrument that tests whether fusing volume, sentiment and technical signals holds up under purged cross-validation and realistic costs. Edge on real data is unproven.