# Security policy
## Reporting a vulnerability
Report privately through GitHub:
**[Open a security advisory](https://github.com/micheltlutz/molt/security/advisories/new)**
Please do not open a public issue for a vulnerability.
Include what you have: the input, the rendered output you expected, the output you got, and
the crate version. A failing test is the fastest possible report.
You should get an acknowledgement within a few days. If a fix is warranted it ships in a
patch release, and the advisory is published once users have had a chance to upgrade.
## Supported versions
The latest released version. There are no long-term support branches. When 1.0.0 ships, 0.1.x
stops receiving fixes: [`MIGRATION.md`](MIGRATION.md) is the way forward, and several of the
fixes below changed output, so they are not backported.
## How the guarantees map onto the OWASP Top 10 (2021)
Molt generates strings and writes files, so most categories do not apply. The ones that do,
and where each is enforced (the full table of inputs is in
`.claude/skills/molt/references/security-owasp.md`):
| A01 Broken Access Control | path traversal and destructive operations in `StaticSiteGenerator` | component whitelist, symlink refusal, a pure `clean()` guard |
| A03 Injection | text, attribute values, names, comments, URLs, JSON and XML reaching the output | escaping on entry, write-time validation, the URL policy, `escape_json_script`, `escape_xml` |
| A04 Insecure Design | an unsafe call that is only discouraged in a comment | types that cannot be built wrongly (`Markup`, `ChangeFreq`, private `indent`) |
| A05 Security Misconfiguration | CI and release | pinned actions, read-only token, a release gated on `main` and green CI |
| A06 Vulnerable Components | the dependency graph | `cargo deny`, `Cargo.lock` committed, optional runtime dependencies only |
| A08 Integrity Failures | half-written files, the release pipeline | atomic writes with unique temporaries; verify and publish in separate jobs |
| A09 Logging Failures | not applicable: the library does not log | — |
| A02, A07, A10 | no cryptography, authentication or network access | — |
## What counts as a vulnerability here
This crate's security surface is narrow and specific: **it generates HTML strings, and its
one safety guarantee is that text and attribute values you pass in cannot become markup.**
These are vulnerabilities:
- Any input to `text()`, `Attribute::new`, `Element::attr`, or the escaping functions that
survives into the output as executable markup or as an attribute that breaks out of its
quoting. An attribute name containing a space, quote, `=`, `<`, `>`, `/` or a control
character is omitted rather than written.
- A class, id, style or `data-*` value that can inject an additional attribute.
- An attribute value that ends a raw-text ancestor (`noscript`, `iframe`, `xmp`, `noembed`,
`noframes`, `script`, `style`, `plaintext`) early. `escape_attribute` leaves `<` and `>`
alone, since neither can end a quoted value, but the renderer writes them as `<` and
`>` for every attribute under one of those ancestors; the fixtures are unaffected.
- `Node::comment` content that can terminate the comment early.
- A tag name that survives into the output with a space, quote, `<`, `>` or `/`. A tag
that is not a letter followed by letters, digits, `-`, `_`, `.` or `:` is **not
rendered, children included** — see `molt::core::tags::is_valid_tag_name`. This covers
`Element::new`, `new WElement(tag)` and `element(tag)` in WASM.
- `StaticSiteGenerator` writing outside its output directory (a `..` component, a root or
drive prefix, or an existing symbolic link on the way), or `clean()` deleting outside it.
`clean()` refuses `.`, the working directory and its ancestors, the home directory, and
top-level directories such as `/usr`. A symlink created *while* a build runs is a race this
crate does not close.
### Rendering depth is no longer bounded by the stack
The writer walks an explicit work stack instead of recursing, and `Element` tears its
subtree down the same way, so nesting depth costs heap — bounded by a tree that already
fits in memory — rather than stack. A 100,000-level tree renders and is freed in the test
suite, in both render modes.
Before this, a deep enough tree exhausted the stack, and a stack overflow is a process
**abort**, not a catchable panic — `catch_unwind` would not have saved you. That made
nesting depth a denial-of-service vector anywhere it could be influenced by untrusted
input. Fixed before the first release under the name Molt.
**What deep trees still cost is output size.** Pretty mode writes one indent string per
level on every line, so output grows with the square of the depth: a 100,000-level tree
pretty-prints to tens of gigabytes. If depth comes from untrusted input, bound it, or
render compact, set an empty indent with
`RenderOptions::pretty().with_indent("")`, or bound it with
`RenderOptions::with_max_depth(512)`, which drops elements nested deeper than that (browsers
flatten beyond roughly 512 anyway). The indent is also restricted: `with_indent` keeps only
spaces and tabs, at most 16, because it is written at the start of every line and a string
with markup in it (`renderWithIndent(untrusted)` from JavaScript) would put that markup on
every line.
`Document::render`, `audit`, `Clone` and `PartialEq` are iterative as well, so a document, an
accessibility audit, a copy or a comparison of a 100,000-level tree is safe. **`Debug` is
not**: `{:?}` (and the failure message of `assert_eq!`) on a tree that deep recurses and
aborts. Compare with `assert!(a == b)` in tests that build trees that deep.
One residual, and it is not a stack limit on rendering: a bare `Node::Fragment` chain with
no element anywhere in it is still freed recursively. Giving `Node` its own `Drop` would
forbid `match node { Node::Element(element) => element }` — a partial move this crate's own
`static_site` example performs — so the iterative teardown lives on `Element`. Every
fragment reachable through an element, which is every fragment in a document, is freed
iteratively.
These are **not** vulnerabilities, because they are documented behaviour:
- `raw_text()` and `Node::Raw` not escaping their input. That is their entire purpose; they
are named and documented so that an audit can grep for them.
- `<script>` and `<style>` content not being escaped by default. For
JSON data in a `<script>` (JSON-LD, a data block) use `seo::json_ld` or `seo::script_json`,
which write `<`, `>` and `&` as JSON escapes so a `</script>` in a value cannot end the
element. Untrusted text in script *code* has no safe form; do not put it there.
- `Attribute::raw` not escaping. It exists for library-controlled keys. It does **not**
skip the URL check below: a raw `href="javascript:..."` is still rendered as `href="#"`.
### Sitemap and feed output
`escape_xml` drops the characters XML 1.0 cannot represent (U+0000–U+0008, U+000B, U+000C,
U+000E–U+001F, U+FFFE, U+FFFF), because one of them makes the whole document ill-formed and a
reader then rejects every entry. A sitemap `<loc>` and an RSS `<link>` that are not absolute
`http(s)` URLs are left out, since feed readers and crawlers present them as something to
open; `isPermaLink="true"` is written only for an `http(s)` guid. The sitemap protocol's
50,000-URL and 50 MiB limits are exposed as `MAX_URLS`/`MAX_BYTES` with
`SitemapGenerator::chunks`, not enforced silently.
### URL attributes are checked at write time
`href`, `src`, `action`, `formaction`, `poster`, `cite`, `background`, `xlink:href` and
`data` are rendered as `#` when the value, after one character-reference decode and after
dropping characters at or below U+0020, starts with `javascript:`, `vbscript:` or
`data:text/html` (ASCII case-insensitive). `data:image/...` is left alone, and so is a
non-URL attribute such as `title`. Comparison is on the stored value, so a value that went
through `Attribute::new` has already turned `&` into `&` and is not decoded back into a
scheme.
Also checked: `ping`, `manifest` and `longdesc`, and `srcset`, whose comma-separated
candidates are each checked (a `javascript:` URL in any one blanks the attribute). On
`iframe`, `frame`, `embed` and `object`, a `src` or `data` that starts with `data:` is
blocked unless it is a raster image (`png`, `jpeg`, `gif`, `webp`, `avif`, `bmp`): those
elements load a document, and `data:image/svg+xml` or `data:application/xhtml+xml` run script
in it. The same SVG as an `<img>` source is inert and is left alone. `srcdoc` is not a URL but
a whole HTML document the frame runs; it is written empty for any non-empty value.
`Attribute::trusted_url` and `Element::trusted_url` (and `trustedUrl` on the WASM element)
skip that check, `srcdoc` included. The value is still attribute-escaped. Use them only for a
URL, or a `srcdoc` document, you wrote.
Not checked: `<meta http-equiv="refresh" content="0;url=javascript:…">`. The HTML
specification tells the browser to ignore a `javascript:` URL there, and checking it would
need the element's other attributes. Event-handler attributes (`on*`) and `style` set through
`attr()` are not filtered either: do not map untrusted *keys* into `attr()`.
The design, and the work that is deliberately not in this crate, is in
[docs/plans/2026-10-17-seguranca-e-performance.md](docs/plans/2026-10-17-seguranca-e-performance.md).
## CI and release
What the workflows do, and what only the repository owner can switch on.
- Every third-party action is pinned to a commit SHA with its version in a comment, and
Dependabot's `github-actions` ecosystem keeps the pins current. A tag such as `@v3` can be
moved by whoever controls the action's repository; a SHA cannot.
- `ci.yml` is `contents: read`. Jobs have a `timeout-minutes`. Checkouts use
`persist-credentials: false`, so the token is not left in `.git/config` for later steps.
- The release workflow refuses a tag whose commit is not on `main`, or on which the CI
workflow has not already succeeded. The publish job names the `release` environment.
- `scripts/verify.sh` writes its log to a `mktemp` file, not a fixed path in `/tmp`.
**Owner steps, not code:** in the repository settings, add required reviewers to the
`release` environment; and register `molt-framework` as a Trusted Publisher on crates.io so
the long-lived `CARGO_REGISTRY_TOKEN` secret can be retired (the workflow still uses the
token until then, so a release keeps working).
## For users
- Pass untrusted data through `text()` and `attr()`, never `raw_text()` or `Node::Raw`.
- If you must inject markup you did not author, sanitize it with a real HTML sanitizer
first — this crate escapes, it does not sanitize.
- Run `molt::accessibility::audit` in development; several of its findings are also
signs of markup being assembled by hand.