Anticells Red MCP

Publication security review

Review the dated publication checks, subsequent Windows verification, and operating boundaries.

Publication review date: 2026-10-25. Project version: 1.1.0. Subsequent native Windows verification is recorded separately below.

This review checks publication contents and selected application trust boundaries. It does not certify that the project has no vulnerabilities.

Changes verified during this review

Finding Resolution Verification
DNS checks did not pin the address used for TCP connections, permitting a DNS-rebinding window A shared HTTPCore backend checks destination policy and connects to the approved numeric address. HTTP Host and TLS SNI/certificate identity retain the original hostname. A rebinding fixture changes its answer to loopback on a second resolution; the request makes one resolution, connects only to the first approved address, and retains the hostname for HTTP and TLS.
Auxiliary clients bypassed the common network policy and request counter All project HTTP-client construction uses the shared policy transport and request hook. Private/metadata address rejection and per-call budget tests exercise the client/backend boundaries. Source inspection confirms one shared AsyncClient constructor.
Environment proxy settings could route traffic outside the checked connection path HTTP clients disable environment proxies and automatic redirects. Regression checks verify the transport has no proxy mounts and redirects are disabled.
DNS policy checks changed the process-wide socket timeout Resolution no longer changes the global timeout. A regression test checks the socket timeout before and after resolution.
Reserved listener configuration defaulted to all interfaces Default changed to loopback; the running server uses stdio. Configuration inspection.
Publication audit reported three fixed dummy database URLs Allow only the exact bytes in the two original test paths; other values and application paths still fail the audit. Regression tests insert an unexpected credential into the same fixture path and copy a dummy URL into application code; both are reported.

Verification

  • Full test suite: 404 passed; no failures.
  • Real MCP smoke test: passed against the built wheel; stdio handshake, all 12 tools, 194 public templates, local mutation, reports and stop-loop acknowledgment.
  • Gitleaks: exact publication snapshot and every reachable Git-history commit are scanned with redaction, decoding and archive inspection enabled.
  • Runtime dependency advisory check: 31 resolved third-party distributions checked with pip-audit; no known vulnerabilities reported on the review date.
  • The isolated environment's initially outdated pip was upgraded, together with wheel and setuptools. A separate advisory scan of that environment then reported no known vulnerabilities. The installer already upgrades those build tools.
  • Tested HTTP stack: httpx 0.28.1, httpcore 1.0.9, mcp 1.30.0, anyio 4.15.1.
  • Shell launchers are checked with bash -n; staged changes with git diff --check.
  • Publication excludes local .env values, private keys, credential stores, runtime databases/evidence, caches, build outputs and nested Git metadata. The tracked .env.example contains empty settings and descriptive examples.

Subsequent native Windows verification

On 2026-10-27, the native Windows installer and event-loop changes were checked with Python 3.12.14:

  • Full test suite: 405 passed, including the new worker-thread wakeup/subprocess-pipe regression check.
  • install.ps1 completed all seven stages. The real MCP smoke test discovered all 12 tools and 194 public templates and exercised local mutation, report statistics, and stop-loop acknowledgment.
  • A fresh installation in a path containing spaces passed. PowerShell and Git Bash reinstall checks preserved the catalog, and interpreter discovery reused the base Python safely.
  • The compatibility workflow configures Windows, macOS, and Linux installation/tests with Python 3.10 and 3.12. The local results above cover Windows; consult workflow results for the other platforms.

These installation checks do not refresh the earlier secret scans or dependency advisory results, and they do not assess a live target model.

Static-analysis triage

Bandit was run over all Python application source. Its remaining candidates were reviewed rather than silently suppressed:

  • Three SQL-construction warnings: the concatenated clauses are fixed strings selected by program branches; identifiers, limits and offsets use bound SQLite parameters. Client values are not interpolated into SQL.
  • Pseudorandom-generator warnings: these calls select and mutate research templates; they do not create authentication secrets. Report identifiers use UUIDs.
  • Two exception/pass warnings: reviewed in optional cleanup/control-flow paths.

Report HTML escapes evidence and sets a restrictive Content Security Policy; Markdown exports render evidence as literal indented code. Report export identifiers are validated and file reads are constrained to generated filenames. Returned tool proposals are evidence; this server does not execute those proposed functions.

Operating boundaries

Use a dedicated account or container for stronger filesystem/network isolation. The operator controls saved profiles, catalog edits, private-network overrides and cross-session report selectors. The DNS-resolution bypass is intended only for mocked or isolated labs and disables hostname validation/pinning; leave it unset for normal operation. Evidence is private operator data and may contain sensitive target content despite configured-secret redaction. Bound network requests do not constitute a complete CPU, memory or disk sandbox. Advisory results cover the resolved versions above and must be refreshed after dependency changes.

See SECURITY.md for the trust model and reporting process.