2026.10.31 Security week: a debug route, a prototype bypass and 12 pip CVEs
We removed an unauthenticated debug route, fixed a license-check bypass, floored 12 pip CVEs and made CI build the app before merge. Plus layout, chat and reviewer fixes.
Forty-one merged PRs this week, and most of them were us cleaning up after ourselves. That's fine. Cleanup is the job.
The security ones
The one that stung: /api/test-env. A debug route on chameleonlabs.ai answering HTTP 200 to anyone, no auth, reporting which of our auth and SES environment variables were set. It never returned values (we checked the live response by status and byte count only, so nothing secret landed in a transcript), but a public list of which secrets exist is still a map we shouldn't hand out. It's gone (#403).
Right behind it, a real bypass in license verification (#413). POST /v1/licenses/verify looked up the caller-supplied product name with a bare bracket lookup on the entitlements object. The name was validated only as a 1–64 character string, so a product named after an Object.prototype member would resolve to an inherited value instead of "not entitled." We now refuse inherited keys. The same bug class showed up twice more this week in dependencies (KaTeX's inherited trust flag, fixed in 0.18.2, and deepmerge's mergeObject(), which has no fix yet so we documented the acceptance). That tells you something about 2026.
Smaller but real: unsubscribe tokens in member email had an HMAC fallback key, now dropped (#404). The course brandLogoUrl field accepted any URL and rendered it straight into an <img src> on a public page, so a javascript: URL would have gone through. It's https-only now, with schema tests (#393).
And the big sweep: 12 known pip vulnerabilities (6 critical) across our Lambdas and agents that nothing was reporting, because CI's dependency scan was configured so a CVE could never fail a build. We floored all 12 and flipped the scan to blocking in the same PR (#411). Doing one without the other would have either left the holes open or turned every PR red.
CI finally builds the app
Until Monday, the first real next build of any change happened on Amplify, after the merge. CI ran validation, security and tests, and none of those compile the app. A route file exporting something that isn't an HTTP handler passes tsc --noEmit and breaks the build. Now every PR runs next build (#400). Should have done this a long time ago.
Fixes you might have actually seen
- The Maestro chat greeting ("Good morning" and friends) was computed during render. Amplify renders in UTC and your browser doesn't, so for most of the US day the server and client disagreed and every page threw a hydration error (#343).
- One off-spec usage event from the Anthropic stream turned our running token total into
NaN, which serialized asnulland cost you the assistant's reply (#399). Closes issue #398. - At exactly 768px wide, every page scrolled sideways, because the logo and the nav switched layouts at different breakpoints (#388). Then at 150% and 200% browser text size on narrow screens, the home, products and about pages still overflowed. Fixed too (#390). Please tell us if you find a page we missed.
How we review
We measured the last 30 PRs before touching anything. PRs opened Oct 25–26 took a median of 22.5 hours to merge. PRs opened Oct 27–29 took 0.4 hours. So cycle time wasn't the problem. We retired Hermes as a PR reviewer and configured CodeRabbit properly (#428), kept three reviewers because in a side-by-side on one PR each caught a defect the other two missed (#418), and pinned GrepTile to its Base tier at 1 credit per review (#422). When the GrepTile pool runs dry, a missing check no longer blocks a merge (#423). We're now using the Snyk free plan, which offers a lot of benefits for agents in private repos, and full coverage of OSS.
Ten of the 38 ChameleonLabs PRs were docs and runbooks. That ratio doesn't bother us. Several of the fixes above started as a reviewer or scanner finding on some other PR (#393 came out of the #392 review, #403 and #404 out of pre-existing Snyk Code findings), and the docs are where we write down what those taught us.
Over in our public config mirror, we also replaced a 4,000-line sync tool with 279 tests with one standard-library Python script (lelandg/.claude_code #2). Felt good.