CloudSec.ai — Nate Lee on third-party risk & AI security

Field Notes — Latest

AI export controls hurt defenders, not attackers

Last month I signed Alex Stamos' open letter against the export controls on Anthropic's models. The argument was simple: restrictions on model access hurt defenders and do nothing to attackers. The argument was simple: restrictions on model access hurt defenders and do nothing to attackers. This week

More posts

Use model guardrails to buy defenders time

Use model guardrails to buy defenders time

Love this as in idea to buy defenders time, particularly how they use the models' guardrails to stop an attack. It's basically a prompt injection attack as defense. Embed messages in the tokens saying you're trying to create a new super-Ebola or that you

Yes, some people are way ahead of you on AI. It matters less than you think.

Yes, some people are way ahead of you on AI. It matters less than you think.

Yes, some people are way ahead of you on AI. It matters less than you think. The field moves so fast that head starts deprecate quickly. The people who got in early spent their first 18 months fighting model limitations that no longer exist and discovering dead ends that are

If it's in the context window, it's an instruction.

If it's in the context window, it's an instruction.

If it's in the context window, it's an instruction. A recent prompt injection POC drives this home with a pull request that adds a config file pointing to a PNG file. A coding agent reads that file, opens the image, and follows the instructions rendered in

Sam Altman called this the fast fashion era of SaaS. He's more right than he intended.

Sam Altman called this the fast fashion era of SaaS. He's more right than he intended.

Sam Altman called this the fast fashion era of SaaS. He's more right than he intended. Good. That will drag software prices down across the board. But there's a difference between using that leverage and living in it. When an engineer one-shots 80% of a

Agentic security reviews must assess the workflow

Agentic security reviews must assess the workflow

The problem with treating agentic security reviews like any other SaaS review is that you never evaluate the thing that actually runs. The usual questions still matter: Is the vendor trustworthy? Do they align with security best practices? And of course... Do they have a SOC 2? Those are not

We just passed 1,000 subscribers on YouTube for Threats, Pitfalls & Risk Myths, The TPRM Podcast!

We just passed 1,000 subscribers on YouTube for Threats, Pitfalls & Risk Myths, The TPRM Podcast!

We just passed 1,000 subscribers on YouTube for Threats, Pitfalls & Risk Myths, The TPRM Podcast! Thank you to everyone who has watched, subscribed, shared, commented, or told someone else to check it out. Haven’t caught it yet? We’re 17 episodes in, and the guest lineup has

AI won’t turn security into a headcount exercise

AI won’t turn security into a headcount exercise

Most board's first instinct with AI is "Great! So we can cut 20% of our headcount?" I chat with Alex Stamos, former CISO at Facebook and Yahoo who currently runs product at Corridor about how reducing headcount has a longer history than people realize and takes

Thank you to the Cloud Security Alliance for the 2025 Juanita Koilpillai Service Award!

Thank you to the Cloud Security Alliance for the 2025 Juanita Koilpillai Service Award!

Thank you to the Cloud Security Alliance for the 2025 Juanita Koilpillai Service Award! I wrote two papers on building secure agentic systems - one waaaay back in 2024 with Laura Cristiana Voicu and another in 2025 with Ken Huang. At the time, it didn't seem all that significant

Next year some companies will be confused about why they're slow.

Next year some companies will be confused about why they're slow.

Next year some companies will be confused about why they're slow. Not demoralized. Not obviously broken. Competitors will be shipping faster, thinking faster, moving faster and there won't be a clean explanation on any dashboard. Atlassian's new research made me 🤨 with it's