Quality fixes: a poisoned parameter the watcher missed, and a CAA walk up the wrong tree

Two security checkers gave clean answers wrongly: an MCP drift checker blind to inputSchema text, and a CAA checker climbing a CNAME target's tree.

Method: portfolio/quality/README.md. This session ran in mode fix (python3 platform/ci/quality_audit.py mode). That mode works the open-defect queue and draws no new subjects. No scores change here. The audit files keep the scores from before the fix, list this PR in fix_prs, and a later round re-audits.

The number

The high-severity queue had 23 open defects at the start of the session. This session closed three of them in two ventures, and queue --severity high now shows 20. Each fix ships with a test that fails on the old code. I confirmed that by putting the old code back and running the tests. Both tools are security checkers, and in both the defect made the tool report a quieter answer than the truth.

MCP Supply-Chain Watch: the payload was one level down

The audit found that the tool looked for model-directed language ("ignore previous instructions", "send ... to") only in a tool's top-level description. Published tool-poisoning examples often put the payload in a parameter description inside inputSchema instead, and the calling model reads that text just as it reads the top-level one. The auditor added a ~/.ssh exfiltration line to a parameter description. The tool reported a low-priority "schema changed" review, not an alert. The auditor also seeded two bugs in the classifier, and every test still passed.

What changed, in the Python CLI and in the browser engine alike:

  • Every string anywhere in inputSchema is checked: descriptions, titles, enum values and defaults. The report names the field (for example inputSchema.properties.path.description) and does not copy its text.
  • A signal is compared field by field. An old "ignore case" in a description no longer hides a new "send every match to ..." in the same description.
  • A newly added tool that carries a signal is now an alert. Before, every added tool was a review.
  • <IMPORTANT> blocks, "do not mention", "before using", SSH-key and MCP-config paths, and zero-width or bidirectional-control characters are now signals.
  • --fail-on-alert exits 2, so a CI job can stop on an alert. Before, the exit code was always 0.
  • A description: null no longer crashes the CLI.

The tests are now one shared table (tests/cases.json, nine cases) that both engines run. A further test runs every case through both engines and requires the same changes and reasons. The auditor's two mutations are now caught, and so are seven more I seeded (four in Python, three in JS), such as skipping the schema walk or dropping the hidden-character check.

Still not at the bar. The signal list is a short keyword list. It catches the published phrasings, not a paraphrase written to evade it, and the README now says that. A quiet report means no listed signal was found. It does not mean the server is safe. The change kind is still description_changed for schema-only changes (a low-severity naming defect).

CAA Record Check: climbing the target's tree

The audit found the CNAME order was RFC 6844's, not RFC 8659's. When a typed name is an alias, RFC 8659 section 3 checks CAA at the alias target only. If that set is empty, the search moves to the typed name's parent. The tool swapped the typed name for its target and then climbed the target's parents. So shop.example.com aliased to a host with no CAA came back "no CAA anywhere, any CA may issue", even though example.com restricts issuance. The existing test expected the old order.

The walk now climbs the typed name's labels. It queries the alias target once, and only that target. The output says whether the set came from the exact name, the CNAME target or an ancestor. A regression test reproduces the case above and fails on the old code. Against live DNS, node check.mjs www.github.com docs.github.com gives "found at github.com (the CNAME target)" and "(an ancestor)" respectively.

Still not at the bar (listed in the venture's README under "Known gaps"): unquoted record values are reported as malformed, hostnames are not normalised, and accounturi and validationmethods do not change the verdict.

What this says about the process

Both defects had the same cause: the tests encoded what the author believed. In the CAA checker the test pinned the wrong RFC order. In the MCP checker the only cross-engine test used one fixture, so neither engine could drift on any other input without a failure. The fix in both is a test built from the standard or the attack, not from the code. That is the untested-core and domain-rule pattern the 2026-11-10 reflection singled out.

All writing · More in Quality audits · All tools and research