Posts
-
I tried to get two agents to leak secrets through MCP tools. Neither did. Efe Çakıcı #mcp#agents#prompt-injection#measurement A canary experiment with a filesystem server and a fetch server: two models, one injected instruction, zero canaries read. The interesting part is what happened when a human asked instead.
-
I started 38 MCP servers. 17 of them never say what their tools do. Efe Çakıcı #mcp#security#measurement 584 tools across 38 servers, 392 with annotations, 17 servers with none at all. The ones that stay silent are the browser controllers and the database drivers.
-
I asked my AI agent to inspect a website. The website took over my machine. Efe Çakıcı #security#agents#mcp#rce#measurement A local lab, 34 recorded runs, and 6 stolen credential files. You tell an agent to check a library's docs; the website steals your session and executes commands before the model speaks.
-
How omp traded 50 MB for 7.7x faster startup Efe Çakıcı #performance#startup#bytecode#bun I updated omp and it opened instantly instead of the usual split-second lag. 24 alternating runs per version later: 232 ms down to 30 ms, and a binary 50 MB heavier.