I build AI agents and test them for made-up answers and tricks that leak data.
I build AI agents that do real work, with the security inside: tools that reach only what they need, your OK before risky actions, and tests you can rerun. I also test the harnesses agents run on, and publish every finding.
Research
All researchIndependent tests of AI agent harnesses: who the model listens to, which attacks get through, which defences hold.
Post · 17 Nov 2026
Approval gates for AI agent tool calls, tested in 1,940 runs
Three rules for when an AI agent must ask before it acts, tested on an email agent with planted attacks: what each one stopped, and how often it asked.
Post · 16 Nov 2026
Who does an AI agent obey? Three harnesses tested
Operator, user, tool or background job: which one Claude and GPT follow, and how much the agent harness changes it. Measured through three harnesses.
Post · 16 Nov 2026
Prompt injection on an email agent: 960 runs
An email agent with a password file and planted emails: what leaked, and why "data, not instructions" tags stopped Claude doing its job.
Post · 26 Oct 2026
Pydantic AI vs a hand-rolled agent loop, measured in 285 runs
One email agent, built by hand and on Pydantic AI, on Claude Haiku 5.5. Where the framework helped, and an output-mode gap that vanished on a rerun.
Writing
All writingSeries to read in order, and posts from real builds.
Series · 8 parts
Testing AI agent security
How agent harnesses decide who each message speaks for, and what that does to prompt injection, measured on real APIs and a real email agent.
Series · 6 parts
How an LLM runs
What happens inside a language model when it answers you, one step at a time, measured on a real model.
Series · 2 parts
MCP in practice
How MCP works, what I learned building and using MCP servers, and how local and remote servers connect, step by step.
Post · 21 Oct 2026
Re-checking 413 web pages with Jev: 23 needed a full read
How affiliateprogramterms.com keeps its terms current: one Jev yes/no request per page, $0.04 a refresh, and a test against Claude Haiku and Opus on 66 pages.
Post · 15 Oct 2026
Jev 1.14 vs OpenCV YOLOX: image questions and bounding boxes
I tested Jev 1.14's image input and new Box answers on 6 photos and compared the boxes with OpenCV's YOLOX detector. Accuracy, box fit, speed and cost.
Post · 3 Oct 2026
How I calibrated an LLM judge to grade like me, 25× cheaper
Evals on 188 answers from ChatGPT, Claude, Chatbase and my own pipeline over product datasheets, and a hybrid judge with Jev that grades like I do.
Post · 2 Oct 2026
Can an AI agent submit Upwork proposals? Yes, with your OK
Upwork's official MCP lets Claude, Codex or Cursor send a proposal for you. What it shows before sending, what stays on upwork.com, and what I got wrong.
Projects
All projectsOpen source, simulations to play, live tools, a model and benchmarks.
Open source
firstwatch
Your agent reads web pages, files and tool output, and anyone can hide instructions there. firstwatch scans for them before the agent acts, and makes it ask you first.
Simulation
The Butler
An AI butler tidies an inbox with real tool calls. Attack it with prompt injection, defend it with rules, locks and approvals, then read the score and the trace.
Live tool
Play with Jev
Games and toys built on Jev, an AI that answers in probabilities, not words. Race it to find Pip, ask the Magic Jev Ball, or let waif name a feeling.
Model
tinytowns
A very small AI model that makes up town names. Type some letters and see its guesses for the next letter. It does not know which places are real.
Benchmark · 14 Oct 2026
Small model vs LLM
One small job, five kinds of tool: a 4 MB model, open language models, Claude and a plain list. What each one is good at, and what it costs.
Open source
Agent skills
Skills you can drop into your coding agent: find Upwork jobs that fit you, turn an article into a video, free RAM, close a session cleanly. Install one, or all with one command.
Simulation
Plug
An AI agent does one job through MCP servers. Turn a server bad, then connect only what the job needs and lock the rest. See the JSON-RPC and the trace.
Live tool
affterms
Affiliate terms for 471 dev and SaaS tools, each figure sourced and dated. A calibrated model says which pages changed; a person reads them before any figure moves.