Skip to content

v0.8.48 — Reproduced head-to-head vs native MCP gateway

Choose a tag to compare

View all tags
@sattyamjjain sattyamjjain released this 18 Jul 03:31
v0.8.48
fc1859b

[0.8.48] - 2026-07-16 — "Reproduced head-to-head vs native MCP gateway"

Added

  • bench(vs-gateway): reproduced payload-gap head-to-head vs a native MCP gateway.
    Stood up a live Docker MCP Gateway v2.0.1 and pushed the SAME 12 malformed
    tool-call payloads (type-confusion, amount=-1 value-constraint, ghost args, path /
    URL CVE shapes, argument-injection, over-privileged selection) + 3 benign controls
    through BOTH it and airlock. Dated real number (2026-07-16): the gateway forwarded
    12/12 malformed payloads to the backend; airlock blocked 12/12     at the contract
    layer, with 0/3 benign false-positive and p50 ~0.08 ms/decision. The gateway secures
    transport / identity and sandboxes the server (its logs show secret-scanning +
    no-new-privileges) but does not validate the argument contract — a structural
    gap, not a gateway defect    .
    • benchmarks/vs_gateway/ — reusable corpus + live airlock runner + recorded gateway
      fixture; python -m benchmarks.vs_gateway replays the fixture (no Docker needed).
    • benchmarks/vs_gateway/gateway_harness/ — Docker harness (regen.py) to
      re-measure the gateway live; docs/benchmarks/vs-native-mcp-gateway.md documents
      method + provenance (gateway v2.0.1, docker mcp v0.42.1, engine 29.4.3).
    • README: the native-gateway comparison is now a reproduced benched row + a
      measured line in "Where this fits", not a prose assertion; the LlamaFirewall /
      Invariant cells keep their honest scope-claimed, not re-run framing.
Assets 2
Loading