On 2026-11-14 the "Reset saved inputs" button on two of our calculators threw an error for every visitor. The cause is an old HTML quirk. A control inside a <form> becomes a named property of that form, and the property beats the method of the same name. So <button id="reset"> makes form.reset the button, and form.reset() fails with "form.reset is not a function". That run added platform/ci/check_form_clobbering.py, which fails CI when a form holds a control named after a form method that a script on the page calls.
Today's red-team run seeded ten faults into that check. Its tests caught nine. The one that got through broke the default path used when the check runs with no arguments, and no test ran it that way. Then we read the check the way an attacker would and found three ways past it.
Inline handlers were never read. The check gathered calls from <script> blocks and the page's own .js/.mjs files. <button onclick="this.form.reset()"> is neither, and it throws in exactly the same way. Handler attribute values now count as script.
The form attribute was ignored. A control can sit anywhere in the page and join a form with form="f". It is then one of that form's elements and hides its methods just as well. The check only looked between <form> and </form>. It now works out which form owns each control: the form around it, or the one its form attribute names. A form attribute that names no form gives the control no owner.
data-name="reset" failed the build. The pattern \b(?:id|name)= matches after a hyphen, because a hyphen ends a word. A data attribute that names nothing was read as a control called reset. Attributes are now parsed by their full name.
The same \b pattern was in check_pages_links.py, the check that follows every link on the built site. There it did harm both ways. data-id="compare" counted as a target for #compare, so a dead in-page link passed. And a lazy-loading data-src pointing at a file that does not exist yet failed the build. A one-character lookbehind, (?<![\w-]), fixes both. The new test is split into two assertions for a reason. Our first version checked the total count of problems, and it passed before the fix because the two errors cancelled out: one missed problem plus one false one still made one.
None of the three holes had let a real bug onto the site. Searching every venture page found no handler-attribute call and no form=-joined control named after a method. So this run closed gaps before anything slipped through them. The run record is prompts/runs/2026-11-17-session_01E5sJKik6BThGQESdSEJxXH.json: 18 faults seeded across two targets, 16 caught before the new tests and 18 after, and four real defects in checks, all patched.