Michael
WilliamsLeveraging AI. As it evolves, so do I.
Cybersecurity Engineer II at SEARHC. Cybersecurity engineering, AI security, AI and software risk in a healthcare environment. Five years of security operations, and at every hop I stepped into the unknown and built things: solutions, my teams, and myself.
It started in a crisis.
Distributed teams, regulated data, zero tolerance for downtime. I learned the controls by standing them up under pressure: provisioning, authentication, email security, access lifecycle.
Then I credentialed the gaps.
The degree gave me the map. The SANS scholarship and the certification run gave me the reps on paper: GSEC, GCIH, GCIA, SecurityX, and a dozen more. ChatGPT arrived and became a mentor, thought partner, and assistant that stretched my capabilities.
AI THREAD · EARLY ADOPTIONIt was also my first AI security lesson, learned by accident: the model was confidently wrong often. Meaning, there were gaps, and where there are gaps there are vulnerabilities. Follow the evidence started here.
I almost started at the finish line.
BitDefender, Wazuh, Splunk, TheHive, Action1. I wrote the SOPs, playbooks, detection runbooks, and IR procedures, then watched a team of 30 run them. The architecture worked. The gaps were mine, in the hands-on reps underneath it.
AI THREAD · CURIOSITYEvery playbook I wrote was a workflow a model could one day run. I started asking what would have to be true for that to be safe. That question became the next three years.
3,000 endpoints. 60 clients.
Tier 2/3 triage, hunts that found what the automated layers missed, EDR and SIEM rules tuned across tenants. I built ThreatPrism inside Swimlane SOAR to extract IOCs, map ATT&CK, and hand analysts context before they touched the alert.
AI THREAD · STUDY, THEN APPLIEDPostgraduate AI/ML at UT Austin McCombs. Notebooks taught me precision versus recall the hard way. ThreatPrism was the first time a model touched production work, and the first rule I wrote for myself: the model adds context, a deterministic gate makes the call.
If it can be prompted, it can be attacked.
I pointed the detection mindset at the models. Red-teamed chatbots and agents against OWASP LLM Top 10 and MITRE ATLAS, built guardrails that run before and after the model, and fine-tuned one that refuses to become a shell.
AI THREAD · ADVERSARIALThis is where curiosity turned into a practice: threat model the model, write the should-fail case first, and never let the model gate its own security decision.
Agents will transact. Who vouches for them?
By day: a greenfield cybersecurity program. Detection engineering, vulnerability testing and remediation, threat hunting, AI and software risk, operations engineering across email, endpoint, network, and IAM lifecycle, and agentic AI security with Copilot in a healthcare environment.
By build: AgentLedger, a six-layer proof of concept for discovery, identity, trust signals, context disclosure, workflow validation, and liability evidence. Infrastructure, not an orchestrator. Docs · Repo
AI THREAD · BUILDING WITH IT, FOR ITBy now AI was both the tool and the target. I built with agent frameworks and MCP servers, then turned around and built the controls those agents need: deterministic gates, monitored deployments, one governed instruction set across every runtime.
Everything is untrusted until proven otherwise.
Guardrails are engineering, not an afterthought.
Findings trace to evidence, or they are not findings.
sha256:e2c9…5f10 · this sha256:4a07…b93dthis page is tamper-evident · every rebuild appends to the ledger