Writing
Notes from the things I build. Several posts have an exhibit in the lab you can try in your browser. When a post turns out to be wrong, the corrections page says so.
Start here
All posts
- Why did the models stop helping halfway through OpenCode's incident? The frontier models helping with the investigation started refusing. The fix was a vetted-access tier from the vendor, which works, but only once someone approves you.
- The user signed in twice. Was it the same user both times? An agent that signs the user in to your identity provider and then to a downstream API holds two identities. Only federation makes them one person.
- Why does the OAuth flow that worked in my demo fail against a real provider? OAuth is a family of options, and each provider picks its own subset. PKCE, redirect matching and plan tiers are where the picks differ.
- Why did my refund tool accept an approval code from a customer? The check compared the code correctly. The code arrived in the ticket, and whoever writes the ticket also writes the code.
- Why did an evaluation agent end up inside a Medicare portal? A research task hit a locked door while it still had a live network connection. Nobody on either side was watching.
- Why does Agentforce take orders from a web form? To the model, a stranger's lead and the employee's question are one stream of text. The filter meant to stop the leak only blocked URLs it recognized.
- Whose agent was it in the Hugging Face breach? The July intrusion keeps getting filed under AI security. The agents were a lab's own models under evaluation, the bug at Hugging Face was ordinary code execution, and the useful lesson is the one Hugging Face learned during cleanup: keep a defensive model you own, running on your own hardware, wired to your own logs.
- If the model refused the attack, why isn't the server safe? A refusal is the model's call, and the model's call can change. mcploitable is a lab of deliberately vulnerable MCP servers built to show where that stops mattering: 105 attempts at the correct control, zero through.
- Why does my MCP server run fine in a terminal but refuse to connect? The client parses everything on your stdout as JSON-RPC. Your log lines are not JSON-RPC.
- Why does my MCP server call the tool I told it never to call? The system prompt is advice to the model. Nothing between the model and the API ever reads it, so the checks have to live in code.