
I killed six of my own results in one night. That was the win
I built an AI security benchmark this week. By the end of one night I had killed six of my own...

Author profile
AI security engineer focused on how LLMs fail — prompt injection, jailbreaks, and agent behaviour. I build systems that stress, break, and harden AI under real-world conditions.
Browse the latest writing surfaced through DevArt.

I built an AI security benchmark this week. By the end of one night I had killed six of my own...

My anomaly detector just scored a perfect 1.000 AUC. Caught every bad sample, zero false...

For 4 months I've been building something different: a security AI "council" — specialized red-team,...

I’m not approaching this as a red team or a blue team problem — I’m looking at the entire system. So...

This specific undertaking is not fundamentally burdensome in terms of labor; however, this endeavor...

Every time you send a query to a cloud AI API, you're sending data you don't control. For most use...

AI platforms have a unique failure mode: they can bankrupt you. A runaway inference loop. A...

The Model Context Protocol (MCP) is how AI agents talk to tools. Claude Code, Cursor, Windsurf — they...

Part 3 of the quantization series. Yesterday I tested whether Part 1's drift-inversion intervention...

Part 2 of the quantization series. Spent two days and $12 hunting for the right localizer after Part...

Spent the week reproducing a quantization bug that broke a quiet assumption underneath every...

This is a submission for the Hermes Agent Challenge I've spent the last five months building an...
98% adversarial defense rate. 200 rounds. One $700 GPU. I just finished benchmarking my defender...
This is a submission for the Gemma 4 Challenge: Write About Gemma 4 I've been building an offline,...

5-defender consensus swarm + autohealer hit 100% defense rate by round 400 after only 6 breaches in...

200 adversarial rounds. 4 breaches. 98% defense. Five 1.5–7B local models on a single RTX 5070 beat...

500 adversarial wargame rounds. 5 breaches. 108 auto-blocked. From 73%% to 99%% in five iterations on consumer hardware.

Four iterations of adversarial wargames on consumer hardware. Model swaps, Vanguard prompt injection, and an auto-healing immune system that patches the swarm mid-run.
Advertisement