On this page
Security: how Snaglist protects people, and how it was tested
Snaglist runs inside other people's websites and sends what it finds to GitHub, an email or the site's own server. So it must never carry away what someone entered or who they are, never let the page or a link change what it sends, and never put the site at risk. This page says how it does that, what the security check in November 2026 found and fixed, and which tests pass today.
To report a problem privately, use Report a vulnerability on the repository's Security tab; read SECURITY.md. For what a report holds, read What a report holds, and what it never holds.
Latest results: 8 November 2026.
How Snaglist protects people
What it never takes
- Nothing entered in the page: field values, editable areas and the choices made in lists, tick boxes and radio
buttons are left out. A step says "Entered text in Your name" or "Made a choice in Title", never what it was.
- No cookies, storage, accounts or query strings. Addresses keep their path, with ids replaced by :id; the
query and the hash are taken out, including from addresses inside error and console messages.
- Emails, phone and card numbers, IP addresses, ids, keys and tokens are replaced in everything a report holds,
and anything inside data-snag-private is left out.
- The whole report is shown to the person before it goes, and can take parts out.
What the page cannot do - Snaglist builds everything as text, never as HTML, so text from the page is shown, never run. - Text from the page cannot break out of a report's tables or code blocks, and cannot mention anyone on GitHub. - When Snaglist is removed, it puts back everything it wrapped, but never undoes a wrapper another tool added later.
What it loads - One file, no libraries. The optional axe-core is pinned to one version and checked against its hash.
The security check, November 2026
Snaglist 0.3.1 was checked in three ways: the free scanners in Security scans, an independent code review by a reviewer who had not read the scan results, and new tests in a real browser for every real finding. No high or critical problems were found. 14 medium or low problems were real; all are fixed, and each fix has a test that failed before it and passes now.
| Severity | What was found | How it was fixed | Test |
|---|---|---|---|
| Medium | The set-up tool fills its fields from a shared link. A crafted link could put a shell command into the label commands, code into snaglist-start.js and instructions into the AI prompt. | The repository, the app details value and the language code must have the right shape, or they are left out with an error under the field. The button words are quoted in the prompt. An unknown choice in a link is ignored. | SEC-04 |
| Medium | Recorded steps and the trail kept the choice made in a field: the option chosen in a list, a ticked box, a selected radio button or a picked suggestion. | Only that a choice was made, and in which group or field, is kept. | SEC-05 |
| Medium | Text entered in an editable area (contenteditable) was kept as text on the page, and private parts inside a chosen part were not left out. |
Editable areas and private parts inside the chosen part are left out. | SEC-01 |
| Medium | Error and console messages kept addresses with their query strings, and logged objects with their values. | Each address keeps its path only; a logged object shows its keys only. | SEC-02, SEC-06 |
| Medium | Several key and number formats were not replaced: access_token=, Authorization: Bearer, session=, pwd=, AWS key ids and phone numbers written with dots. |
Each is replaced now. Error codes, timings and version numbers stay. | SEC-07 |
| Medium | A part's HTML kept phone numbers in tel: links, names in skype: links, and values in aria-valuetext, aria-description, id and data-testid. |
A link to anything but a web page keeps only its scheme. Value attributes become …, and every other kept attribute has emails, numbers and keys replaced. |
SEC-08 |
| Medium | The example worker for your own server could be used by anyone to make issues, because the origin check stops other websites but not people. | The example uses a rate limit when one is set up, refuses an oversized request before reading it, and stops mentions. The help says to add a rate limit before it is public. | SEC-13 to SEC-16 |
| Low | A console message or an address could break out of a report's code block, and mention people on GitHub. | Messages and steps stay on one line, and an invisible space follows every @. |
SEC-09 |
| Low | A backslash or \| in page text could add a column to a report's tables (found by CodeQL). |
Backslashes are escaped too. | SEC-19 |
| Low | In WebKit (Safari's engine), cleaning a very long console message took 1.5 seconds, holding up the page. | Messages are cut before they are cleaned, and the email pattern is bounded. | SEC-10 |
| Low | Cutting text in the middle of an emoji stopped GitHub and email sending. | Text is never cut through an emoji. | SEC-11 |
| Low | axe-core was loaded without a hash. | The default copy is checked against its hash; check.integrity gives one for your own copy. destroy() takes it away again. |
SEC-03 |
| Low | destroy() could remove a wrapper another tool added after Snaglist. | It puts back only its own. | SEC-12 |
| Low | The privacy check let a private word through when written with "TRex", and ignored a word starting with a hyphen. | A private word is let through only inside an allowed public name. | SEC-17, SEC-18 |
Also changed
- The review window's privacy statement no longer says no personal information is kept, because the page title
and the text on the page are copied as they are. It says what is left out, and asks people to check the rest.
- The test and screenshot tools serve only the files git would publish, so .git and .private-words are never
on offer, even on this computer.
- tools/check-private.sh stops if it cannot find the repository (ShellCheck).
Left as they are, on purpose
- Page titles and the text on the page are copied as they are, because they are what developers need to find
the problem. They are shown to the person before sending, sites can mark parts data-snag-private, and
page.name can replace the title.
- Names in paths and class names (such as a picture called after a person) are kept, apart from ids,
numbers, emails and keys. They are part of the site's structure.
- A report sent to GitHub travels in the address, so it is in the browser's history. That is how GitHub's
new issue form works; Copy as Markdown, Download as JSON and your own server avoid it.
- Answers are kept in the tab's session storage while a recording runs, so they survive going to another
page. They are cleared when the recording stops, is cancelled or Snaglist is removed.
- The demo and set-up pages on GitHub Pages have no content security policy, frame or content type headers,
because GitHub Pages cannot set them, and they load the design system's style sheets without a hash, because
those change with each design system update. Nothing from a visitor is written into the pages as HTML, and
they hold nothing worth tricking anyone into.
- Snaglist's window is an open shadow root, so the page's own scripts can read what is entered in Snaglist's
form, as they can for any part of the page. A site that records sessions should leave Snaglist out of them.
Security tests and results
tools/test.sh runs every test: the shell tests here, then the browser tests in the Playwright image, in
Chromium, Firefox and WebKit (Safari's engine). Each prints its ID, and the browser when it is not Chromium.
| ID | What is tested | Result |
|---|---|---|
| SEC-01 | Text entered in an editable area is never in a report | Pass |
| SEC-02 | Query strings and hashes inside console and error messages are taken out | Pass |
| SEC-03 | axe-core is loaded only with its integrity hash | Pass |
| SEC-04 | A shared set-up link cannot put code or commands into what the tool writes | Pass |
| SEC-05 | Recorded steps and the trail never keep the choice made in a field | Pass |
| SEC-06 | Objects logged to the console show their keys only | Pass |
| SEC-07 | Common key, token and number formats are replaced, and error codes, timings and versions stay | Pass |
| SEC-08 | A part's HTML holds no phone numbers, chat names or values in other attributes | Pass |
| SEC-09 | Messages from the page cannot break out of code blocks or mention people | Pass |
| SEC-10 | A very long console message is handled quickly | Pass |
| SEC-11 | A report cut inside an emoji still opens on GitHub | Pass |
| SEC-12 | destroy() leaves wrappers added later by other tools in place | Pass |
| SEC-13 | The example worker files a report, and refuses other sites | Pass |
| SEC-14 | The example worker cannot be used to mention people on GitHub | Pass |
| SEC-15 | The example worker refuses an oversized request before reading it | Pass |
| SEC-16 | The example worker uses a rate limit when one is set up | Pass |
| SEC-17 | The privacy check catches a private word written with "TRex", and lets the public names through | Pass |
| SEC-18 | The privacy check checks a private word that starts with a hyphen | Pass |
| SEC-19 | A backslash or \| in page text cannot add a column to a report's tables |
Pass |
| SEC-20 | The QR code holds nothing entered in the page, and no email address or key | Pass |
| SEC-21 | Snaglist is styled under a strict Content Security Policy (default-src 'self') |
Pass |
The tests for Show as a QR code run the same way:
| ID | What is tested | Result |
|---|---|---|
| QR-01 | The encoder draws the same codes as the design system's terminal/qr.py (tools/tests/qr-fixtures.py makes them), from version 1 to 40 |
Pass |
| QR-02 | The code holds the same address as Open in GitHub, dark on light with a 4-module quiet zone and a description | Pass |
| QR-03 | Offline, Show as a QR code comes first and says why | Pass |
| QR-04 | A long report goes in shorter, the view says what the code holds, and the link and Download as JSON stay | Pass |
| QR-05 | The view works from the keyboard and is dark on light in all four themes | Pass |
The tests for messages, the renamed report types and the design system's rules run the same way:
| ID | What is tested | Result |
|---|---|---|
| MSG-01 | A report that is not sent says what to do in an alert that stays until it is closed, with the raw error under Technical details | Pass |
| MSG-02 | The status region is in the page before the first message, and a message stays while the pointer is on it | Pass |
| TYPE-01 | The report type ids from before 0.5.0 (ux, visual) still work, in settings, text keys and Snaglist.open() |
Pass |
| WORD-01 | The text people read has none of the words the design system's content rules ban (tools/check-wording.py) |
Pass |
| WORD-02 | The wording check catches a banned word | Pass |
| TOKENS-01 | The fallback colours in snaglist.js match the design system's tokens.css, light and dark (tools/check-tokens.py) |
Pass |
On 8 November 2026, TheMightyTRexUK scanned a code from the demo with a real phone, and the phone opened the new GitHub issue, filled in.
The journeys in tools/screenshots.sh go through every main path in a real browser (one part, a recording, a
suggestion, a phone in dark mode, the accessibility check, the demo's display settings, the set-up tool, and
the QR code, offline and in dark mode) and retake the pictures in the help.
Security scans
tools/security-scan.sh runs the free scanners, all in Docker from pinned images. --full adds an active ZAP
scan and CodeQL. ZAP scans a throwaway copy of the pages on a private Docker network with no way out, so it can
never reach other apps on the same computer. Snaglist has no server, sign-in or Docker image, so there is no
signed-in scan, image scan or Hadolint.
| Scanner | What it checks | Found on 3 November 2026 | Now |
|---|---|---|---|
| Gitleaks | Keys and passwords anywhere in the git history | Nothing before the fixes. After them, 2: the made-up keys in the SEC-07 test that checks key formats are replaced | Nothing real. The made-up keys are listed in .gitleaksignore |
| Trivy | Keys, known problems and risky set-up in the files | Nothing (there are no packages to check) | Nothing |
| ESLint, with the no-unsanitized and security rules | HTML built from text, eval, and risky patterns in snaglist.js, the pages, the worker and the tools | No HTML built from text and no eval. 24 general "object injection" warnings and 2 pattern warnings. | Each was read: every key is a fixed name or the site's own setting, so that rule is off; the patterns cannot backtrack and are marked with the reason |
| Semgrep | Semgrep's JavaScript, cross-site scripting and secrets rules, on the same files | Nothing | Nothing |
| Bandit | Risky patterns in the Python tool | Nothing | Nothing |
| ShellCheck | Mistakes in the shell scripts | 1: a cd that could fail silently |
Fixed |
| OWASP ZAP | The demo and set-up pages, baseline and active | No failures. 7 warnings, all about the host's headers, the design system's style sheets and the demo's deliberate ?token=abc request |
Each explained in "Left as they are, on purpose" and marked in tools/zap-rules.tsv |
| CodeQL | GitHub's JavaScript security and quality queries, run here | 4: the table escaping (real, fixed, SEC-19, counted as one finding for its two lines), an unused list (removed), and the demo's Break something button, which fails on purpose | 0 real |
On GitHub, from 3 November 2026: CodeQL code scanning (JavaScript and Python, extended queries) runs on every push, and secret scanning and push protection are on. The first CodeQL scan on GitHub, after the fixes, raised 0 alerts, and secret scanning found nothing. Private vulnerability reporting is on too. Before that, CodeQL was run here with the same engine.
Not tested yet
- Real Safari on a Mac or iPhone, and Edge: the tests use WebKit and Chromium, the engines underneath them.
- The example worker on Cloudflare itself: its tests replace Cloudflare and GitHub with stand-ins.
beforeSend,contextandcomponent.from: these run the site's own code, so what they add is the site's responsibility, and nothing checks it.
Security checks for each release
Every release runs these checks before it is tagged, and adds a row here. A release's notes list its security fixes under Security.
| Release | Tests | Journeys | Security scans | CodeQL | Notes |
|---|---|---|---|---|---|
| 0.5.0 (17 November 2026) | 88 of 88 passed (the browser tests in Chromium, Firefox and WebKit, with TYPE-01 for the old type ids) | Not run: pictures wait until the backlog is done, so the help pictures are from 0.4.0 | Gitleaks 0, Trivy 0, ESLint 0, Semgrep 0, Bandit 0, ShellCheck 0, ZAP no failures (baseline and active) | Run here: 0 alerts | Renamed report types, errors that stay, design system colours; read the 0.5.0 notes |
| 0.4.1 (17 November 2026) | 76 of 76 passed (the browser tests in Chromium, Firefox and WebKit, with the strict Content Security Policy test) | Not run: nothing on screen changed, so no pictures were retaken | Gitleaks 0, Trivy 0, ESLint 0, Semgrep 0, Bandit 0, ShellCheck 0, ZAP no failures (baseline and active) | Run here: 0 alerts | Works under a strict Content Security Policy; read the 0.4.1 notes |
| 0.4.0 (8 November 2026) | 73 of 73 passed (25 IDs; the browser tests in Chromium, Firefox and WebKit) | Passed, pictures retaken, the QR pictures marked up | Gitleaks 0, Trivy 0, ESLint 0 (after one warning in the screenshot tool was fixed), Semgrep 0, Bandit 0, ShellCheck 0, ZAP no failures (baseline and active) | Run here: 0 alerts | Show as a QR code; read the 0.4.0 notes |
| 0.3.2 (3 November 2026) | 55 of 55 passed (19 IDs; the browser tests in Chromium, Firefox and WebKit) | Passed, pictures retaken | Gitleaks 0 (2 made-up test keys listed), Trivy 0, ESLint 0, Semgrep 0, Bandit 0, ShellCheck 0, ZAP no failures (baseline and active) | Run here: 0 real alerts left. On GitHub: 0 alerts | First full check; read the 0.3.2 notes |
This page on GitHub, where the help is written.