Add Sunglasses to Agentic security

[azrollin]

Hi, I am the author of Sunglasses, adding it to the Agentic security table in your format.

It is an open source (MIT) runtime trust scanner for AI agents: 1,089 patterns across 65 attack categories covering prompt injection, tool poisoning, and MCP attacks, plus a mechanism layer that catches attack shapes rather than only known strings. No LLM calls, runs offline. Ships as a pip package, a GitHub Action, and a free web scanner at sunglasses.dev/scan, with a precision/recall benchmark published at sunglasses.dev/benchmark.

Happy to adjust wording or placement.

[wearetyomsmnv]

Nihao, which recall results on your benchmark ?

[azrollin]

Hi! Recall is 97.4% (with 78.7% precision) at v0.3.5, published at https://sunglasses.dev/benchmark

The setup: 38 labeled real agent-input attacks as positives, and 73 real READMEs from famous repos (react, kubernetes, numpy, ollama) as the negative corpus, so precision is measured against real-world false positives, not synthetic text. No LLM judge, deterministic, and the benchmark imports the engine from the local package, so it measures exactly what ships on PyPI.

You can reproduce it yourself:

git clone https://github.com/sunglasses-dev/sunglasses && cd sunglasses && python3 tests/benchmark/precision_recall.py

To be fully transparent: it is our own dataset, which is why we publish the reproduce command so anyone can verify the numbers independently.