Three engagements, written out properly. Clients are anonymized; the numbers are not.
Enterprise vulnerability management · 2026
A 223,596-finding backlog and no process at all
An enterprise cloud estate was generating findings that landed nowhere. No intake, no triage, no risk tiering, no verification, no defined ownership. Cluster resources carried no ownership tags, so a large share of container findings were structurally orphaned — nobody could be assigned them even in principle. There was no governed path to accept risk on findings with no available fix, so those aged in place and inflated the count indefinitely.
The practical consequence for leadership was that posture was unreportable. You could produce a count. You could not answer whether risk was going down, who owned what, or what needed deciding.
I built the process from zero and versioned the baseline, made exploitation rather than CVSS the governing prioritization rule, designed a remediation RACI and secured formal adoption across three teams, and implemented the ownership tagging that made orphaned findings routable at all. The load-bearing move was relocating accountability for unremediated risk from the security team onto the asset owner — which converted remediation from a security request into a business obligation.
100%Baseline dispositioned
129,825Findings remediated
55,532Governed risk acceptances
AI governance in a regulated environment
Letting clinicians use language models without creating a breach
Enterprises were adopting AI faster than they could govern it, and healthcare posed a specific question: how do you let clinicians use large language models against protected health information without creating a HIPAA breach or pulling cardholder data into scope?
Self-hosted models keep data inside the environment but introduce new leak paths — poisoned model weights, prompt and response logs full of PHI, vector stores that bypass record-level access control, and agentic systems that can be steered into unauthorized actions.
Rather than a policy of general principles, I built a hard control: no local model processes PHI until an approval gate is passed and signed. Five mandatory controls, any one of which fails the gate — endpoint authentication, attributable audit, no PHI in logs, retrieval authorization, and model weight provenance. Agentic deployments carry further controls before go-live. Every approved deployment is registered, re-attested annually, and reflected in the HIPAA Security Risk Analysis. Cardholder data never enters the model path at all, enforced through tokenization and a broker pattern.
100%Of models under governance
25%Improvement in threat detection
StablePCI scope, while AI capability expanded
Incident command · PCI environment
The ransomware that arrived through a trusted vendor identity
A user clicked a document link shared from an internal collaboration site. Ransomware followed. The vector is the part worth dwelling on, because it defeats two layers of defense simultaneously: it defeats the user, since every awareness program trains people to scrutinize external mail and this was an internal platform on a domain they used daily; and it defeats the email security stack, because URL reputation and sandbox detonation are calibrated against inbound external mail, not a link to a trusted internal tenant.
Which reframes the investigation. The click was not the incident — it was stage two. The real question was how a malicious document reached an internal library and whose access put it there. The answer was an external vendor identity holding upload and sharing rights far exceeding what the collaboration required, provisioned for a legitimate reason and then never reviewed, scoped down, or expired.
In a cardholder data environment the first question governs everything downstream, so segmentation validation came before anything else. Containment was surgical rather than broad — taking the collaboration platform offline halts the business and creates pressure to rush every subsequent decision. The lessons-learned review targeted the class of problem rather than the symptom: external identity governance, default expiry on guest access, and least-privilege scoping of upload rights.
UnaffectedCardholder data environment
600+Privileged accounts isolated
15Endpoints infected, contained
