
#146 AI Found What Nine Years of Security Research Missed with Charles Guillemet // CTO @ Ledger
Notes de l’épisode
Charles Guillemet started hacking at 12. He sold his piano to buy his first computer. Nine years ago, he built Ledger's offensive security team from scratch; today, he's CTO of a company protecting a fifth of the world's crypto, and he just watched AI find vulnerabilities his team spent almost a decade not finding.
Topics covered:
- The economics of security — and why AI is breaking them
- Ledger's three layers of security, and what The Donjon actually does
- Turning a security problem into a key management problem
- Charles's recommendation for restructuring engineering teams around AI and "harness engineering"
- Zero trust, bus factor, and protecting a company's "crown jewels"
- Using AI to build a threat model and cut noise from production signals
- Why Charles believes the CISO function should stay independent
Timestamps: [00:00:51] Intro & guest welcome [00:01:21] Charles's nerd origin story — hacking at 12 [00:03:33] From software hacking into hardware security [00:04:45] Ledger's security org today: hardware vs. software [00:08:48] The three layers of security at Ledger — and The Donjon [00:13:28] The AI vulnerability that broke 9 years of research [00:14:47] Security as economic asymmetry — and why AI is collapsing it [00:18:31] Why blockchain attackers already spend big to break in [00:20:51] Is prompting AI to "find vulnerabilities" enough? [00:23:48] Security-by-design and turning security into key management [00:26:00] Does AI kill the CTO/developer job — or redefine it? [00:30:40] Rebuilding an SDLC: Charles's small-team, harness-engineering recommendation [00:36:22] Building a threat model with AI [00:38:58] Zero trust, bus factor, protecting the crown jewels [00:42:36] Using AI to cut through noisy production security signals [00:47:04] "Average is better than nothing" — the Wikipedia analogy [00:48:00] Why Charles thinks the CISO function should stay independent [00:49:09] The Easter egg question — advice to his 2017 self
What you'll learn:
- Why the old economics of security are breaking down — and what that means for your own threat model
- A concrete way to turn any security problem into a (much easier) key-management problem
- Charles's recommendation for restructuring engineering teams around small, AI-native pods and "harness engineering"
- What zero trust looks like in practice when developers bring their own devices
- How to use AI to build a threat model and cut through noisy production security signals
Quotes: ⭐ [00:14:22] "The cost of finding vulnerabilities is trending towards zero." ⭐ [00:26:21] "If they think they will continue writing code, yes, they will lose their job." ⭐ [00:23:39] "If you just prompt like crazy — implement this database, implement this feature — and at the end just prompt, 'please find vulnerabilities,' it's something, but it won't solve the security problem." [00:14:47] "Security is a matter of creating an imbalance, an asymmetry between attacks and defense." [00:41:21] "More and more attackers go through the developer door, trying to compromise the endpoint of one developer and then leverage his access to production." [00:47:14] "Some people say AI is only producing average results, but sometimes average is better than nothing." [00:39:19] "If a human has full access to a system he doesn't need, that also means the malware running on his computer has the same access." (All quotes lightly cleaned of stutters/filler — verify against final audio.)
Links:
- Ledger: https://www.ledger.com
- Ledger Donjon (security research team): https://donjon.ledger.com
- Meet Cerberus — Ledger Donjon's AI security harness: https://www.ledger.com/blog-cerberus-donjon-ai-harness
- Charles Guillemet on LinkedIn: https://fr.linkedin.com/in/charles-guillemet
Want to keep talking about this with other CTOs? Join over 600 tech leaders in our Slack community — go to alphalist.com to apply.
Tous les épisodes de alphalist.CTO Podcast – For CTOs and Technical Leaders