Direct Access with Farzan Karimi

In this Direct Access special edition, Community Manager Sean McMillan sits down with Farzan Karimi, Deputy CISO at Moderna, to unpack how his team built an AI threat model agent that cut first-pass security reviews from 2-3 days to about 10-15 minutes. Karimi explains why building the agent was the easy part and how a clean application inventory and a 50-question questionnaire let the team run 687 security reviews in a single month.  

He also covers what happened when the bottleneck shifted from review to remediation, how he separates useful AI security tools from snake oil, and why safety-critical environments like pharmaceutical manufacturing need a more careful approach to AI-driven testing.  

Key takeaways: 

• The data was harder than the AI. Karimi says building the agent wasn't the hard part. The real work was giving it good source material: a reliable inventory of applications and system owners (built partly by tracing contracts back to the services they pay for) and a 50-question, multiple-choice questionnaire that takes owners about 30 minutes. That questionnaire lets the agent run even when no data flow diagram or architecture doc exists. 
• Accuracy was measured, not assumed. Of the tickets the agent created, only 16.7% were canceled because of an agent mistake. The prompt tells the agent not to invent vulnerabilities from missing evidence. Then, a team of about 7 people triages findings and feeds corrections back into the prompt, so each run improves. 
• Automation moved the bottleneck to remediation. Faster reviews meant more risks than application owners could fix. Moderna paused new security engineering reviews for a quarter to focus on triage and remediation support, and leadership approved 7 new positions dedicated to remediation. In this case, AI created jobs. 
• Expect pushback before buy-in. The first wave of tickets drew complaints from developers and the security team for about two weeks. Framing the agent as an enterprise-wide initiative helped and so did showing why a personal chatbot can't run 700 reviews, integrate with Jira, and route findings to the right owners. 
• Agents read what people won't. Because a refrigerator had a registered owner, it went through the agent. The agent read the fridge's brochures and installation guides and found default credentials. The agent also generates interactive threat model diagrams showing how an attacker could move between components. 
• AI without a clear problem is the noise. Karimi is bullish on AI but skeptical of vendors who add "agentic" to a product without solving a real problem. His first question to any pitch: what problem of mine does this solve? Automating a broken process only makes it break faster. 
• In healthcare, safety comes first. Touching manufacturing equipment can trigger a deviation that compromises a production run and, ultimately, patient treatment. Moderna's approach is to pull firmware off critical machinery and test it in an emulated environment instead. 

To hear more from Karimi, watch our session, “Healthcare Security at Machine Speed: $17M Ransomware & AI Adversaries”  on how AI and ransomware are reshaping healthcare security.