IntentProbe — Security skill for Claude Code

IntentProbe

A local scanner for MCP servers, tools, and skills. It reads a frozen model's activations, not just the text — so it catches attacks worded in ways a text classifier never saw.

What it is

IntentProbe runs a tool description or prompt through a frozen local model (Qwen2.5-0.5B), reads a few mid-layers, and scores the mean-pooled activation vector with a small logistic probe (~22 KB). Most scanners read the text itself: patterns, classifiers, rules, or "ask an LLM". This reads the host model's internal state instead.

The point of doing it this way is generalization. When you train a text classifier on attack examples and then face attacks from a source it never saw, the vocabulary often doesn't transfer and recall collapses. The probe holds up better across sources, because it keys off how the mo