Skip to content

GSH v1.1.0 — MCP Tool-Poisoning Playbook and Honest Reference Implementation

Choose a tag to compare

View all tags
@sunilgentyala sunilgentyala released this 15 Jul 21:25
· 21 commits to main since this release
v1.1.0
a594f51

What's in this release

  • Hunt-005: MCP Supply Chain & Tool Poisoning — new playbook covering tool description injection, post-approval schema rug pulls, tool shadowing, and invisible Unicode payloads (approval-time schema hashing, continuous semantic scanning, canary comparison).
  • Sentinel reference implementation (gsh-sentinel-deploy.py, gsh-probe-eval.py) — full baselining, drift-scoring, and ZTLV policy-evaluation logic, runnable end-to-end today.
  • Project website at https://sunilgentyala.github.io/gsh-framework/

Known limitations (read before deploying)

  • The Sentinel scripts ship with a synthetic telemetry generator by default. Every run now prints an explicit SIMULATION MODE warning at startup so this is unambiguous from the console output, not just a code comment. Wiring real LLM gateway or MCP event data into the telemetry-ingestion block (marked # Replace this block in source) is required before this does real detection or enforcement.
  • No packaged CLI, PyPI distribution, or Docker image yet — install via pip install -r requirements.txt and run the scripts directly.
  • No automated test suite or CI beyond Pages deployment yet.
  • No published benchmark numbers (precision/recall/latency) yet — none will be published until measured against real data.

Next milestone

A working MCP proxy adapter (real JSON-RPC interception, policy evaluation, and permit/alert/block against live traffic) is the next planned piece of work — see the repo issues for tracking.

Full Changelog

v1.0.0-beta...v1.1.0

Assets 2
Loading