Writing
Notes on AI security, SOC, and the bridge between them
Working notes from active adversarial testing, defensive operations, and a career split between two security communities that should be talking to each other and mostly aren't.
-
The Defence That Can't See the Attack
I built a prompt injection sandbox with no model in it. The most useful thing it does is show you a control failing for a reason that has nothing to do with how good the control is.
-
Taxonomies That Outlive the Exploit
Every jailbreak list I bookmarked in 2024 is now useless. The categories underneath them are not. Sorting attacks by mechanism instead of by phrasing is the difference between a framework and a changelog.
-
Detecting Prompt Injection: The Playbook Your SIEM Doesn't Have Yet
Prompt injection is succeeding in production right now, and most SOCs can't see it. Why traditional detection primitives partially-fail, three places where they can work, and what's still hard.
-
The Bridge: Why AI Security Needs SOC People, and SOC Needs AI Red Teamers
Two communities that should be talking to each other are mostly not. Here's what each side is missing, and what bridging them looks like in practice.
Currently drafting Threshold tuning for LLM guardrails: where the false-positive cost overtakes the attack cost.
Want a heads-up when I publish a new post? Drop a line through the contact form with "subscribe" in the subject and I'll send a one-line email when something new is up. Or subscribe via RSS. No newsletter platform, no spam, no algorithm.