Skip to content

v1.2.0

Choose a tag to compare

@github-actions github-actions released this 09 Nov 13:48
7c6946d

First release on PyPI: uv tool install agent-perimeter. No change to action.yml's inputs or
outputs; uses: Berakhah/agent-perimeter@v1 picks it up.

Added

  • instrument: bring-your-own-agent injection proof. A loopback-only, scope-gated proxy
    sits between your agent and a server, plants a per-run nonce canary in tool results, and
    records a redacting transcript. A finding is rated by whether your agent repeated the canary
    into a sink and what that sink can reach at call time. It refuses cross-origin, foreign-Host
    and non-JSON requests. See docs/byo-agent.md.
  • Scan notes on the live-scan screen and in SARIF, including why a tool listing was incomplete.
  • examples/scan-reference-server.yml: a copy-paste workflow that scans the MCP reference
    server (@modelcontextprotocol/server-everything@2.1.0) as a stdio target and uploads SARIF
    to code scanning. .github/workflows/examples.yml runs it against every change.
  • .github/workflows/demo-gif.yml records docs/demo/scan.gif from docs/demo/scan.tape on
    a runner with Docker.
  • PyPI metadata (description, classifiers, project URLs) and a PyPI-specific description with
    absolute links.

Fixed

  • Tool listings fail closed. tools/list follows nextCursor with a page cap. A failed or
    malformed first page, a repeated cursor or a page with no tools key is an incomplete
    listing: the scan says so, exits 2 on a gate run, and never writes or compares a drift
    baseline from it. A server that returns zero tools is a complete, empty listing. The
    server-supplied cursor is never echoed into notes.
  • The Action reports drift: unknown, not false, when a gate run was refused before any
    comparison.
  • Injection ratings: only sinks backed by an observed capability rate CRITICAL; a verbatim copy
    of a tool result is LOW data pass-through; description-only sinks are named as such.
  • Instrument redaction: secrets glued to the canary, URL-encoded or nested several rounds deep,
    and values under secret-sounding keys are redacted before anything is written.
  • Reports now carry the installed package version. SARIF tool.driver.version and the scan
    record's tool_version were hardcoded to 0.1.0; they read agent_perimeter.__version__,
    which comes from the package metadata, and a test fails if a literal comes back.
  • --html crashed in any installed copy (the Action, and a PyPI install): importing the report
    renderer read tests/fixtures/corpus.yaml, which only exists in a git checkout. The corpus is
    now read on first use, and a test imports the CLI and renders a report from a copy of the
    package with no tests/ beside it. Found by the new examples workflow.
  • The eval never scores a fixture whose listing stopped early, and the degraded-mode floor test
    can now fail.

Changed

  • Security reports go through GitHub private vulnerability reporting; the personal email
    address is gone and a test guards against a published one. The acknowledgement target is 5
    business days.
  • Every third-party GitHub Action is pinned to a commit SHA.
  • The source distribution contains the package, README, licence and NOTICE only (163 KB, down
    from 2.8 MB); tests, fixtures, the web UI and census data stay in the git repo.
  • Dependencies: patched known vulnerabilities in the Python and web lockfiles (including Mako
    1.4.3).
  • CI no longer uploads the deliberately flawed fixtures' SARIF to this repo's code scanning.

Removed

  • Private project briefs, from the public repo.