mynd parses hostile binary input. Treat parsers, byte and bit readers, decompression, frame composition, metadata, color conversion, allocation, resource limits, dependencies, CI, and releases as security-sensitive.

Version 0.1.0 is a repository foundation and contains no image decoder or encoder. It is not production-ready. Support claims in FORMAT_SUPPORT.md must remain evidence-based and fail closed.

Do not open a public issue containing exploitable details. Use a private GitHub security advisory once repository security channels are enabled, or contact the maintainers privately. Include affected versions, target, feature set, resource limits, a minimal reproducer, and expected impact when safe to do so.

• No panics, hangs, unbounded loops, unchecked arithmetic, or out-of-bounds writes from malformed input.
• Every input-derived allocation and expensive operation is limit-accounted.
• First-party runtime crates use #![forbid(unsafe_code)].
• Published runtime crates use only core, optional alloc/std, and audited first-party mynd-* crates unless an explicit exception is approved.
• Codecs do not access the network, filesystem, environment, clock, or threads.
• Compatibility mode never disables bounds, arithmetic, state, or work checks.
• Errors and logs never include unbounded input-derived text.

The workspace enables overflow checks in its own development, test, and release profiles and uses abort-on-panic for its own release builds. Cargo does not propagate those profile settings when mynd is embedded as a dependency; the downstream application controls its panic and overflow behavior.

Mynd's library guarantee therefore comes from checked arithmetic, panic-free source, forbidden unsafe code, and the lint, test, review, and release gates. It does not depend on a consumer selecting panic = "abort".

scripts/checks.sh
scripts/check_latest_tools.sh
cargo deny check
cargo audit

GitHub CodeQL default setup must be enabled in repository settings. Do not add an advanced CodeQL workflow while default setup is active. See GitHub security settings.

Every release keeps one matching pentest report in security/pentest/vX.Y.Z.md. The report starts before testing, accumulates findings, fixes, CI corrections, and retests, and must end at Status: PASS. There is no separate report-only commit requirement.

After the report reaches PASS, commit the complete candidate and wait for GitHub CI and CodeQL. A failure returns to the same fix, document, commit, and verify loop. The release tag points at the final commit only after its local release gate, GitHub CI, and CodeQL are green.

Required evidence grows with capability: formatting, lints, MSRV/current Rust, feature combinations, target builds, unit and documentation tests, package inspection, dependency policy, truncation/malformed corpora, fuzzing, proofs, Miri, differential/conformance results, and hostile-input regression tests.

Before 1.0, security fixes normally target the latest release line. Once 1.0 is published, this section will list maintained stable and backport lines.