Report privately, not through the public issue tracker:

• Preferred: GitHub's private vulnerability reporting, from the repository's Security tab.
• Alternative: email libraz@libraz.net.

Please include the codehelion version, the platform, the command you ran, and an input that reproduces the problem — or a description of its shape, if the input is a tree you cannot share.

An acknowledgement should arrive within about a week. There is no committed remediation deadline, and no bounty.

Pre-1.0. Fixes land on the latest release only; older versions are not patched.

codehelion reads code that it does not trust — that is the whole job — so the parsing and analysis path is the interesting attack surface. In scope:

• A crafted source file, binary artifact, debug companion, archive, baseline or configuration file that causes a crash, a hang, unbounded memory growth, or reads outside the tree being scanned.
• Any path by which scanning a tree executes code from that tree without the corresponding --allow-execution flag. Nothing in a scanned tree is supposed to run by default, and --untrusted is supposed to permit no execution at all.
• Any network access. codehelion is meant to have none: cargo-deny refuses the common HTTP stacks and clippy.toml disallows sockets in the scan path. A reachable network call is a vulnerability even if it is benign.
• Any path by which source code or analysis results leave the machine.
• Escapes from the compiler-helper process boundary, or a helper response that compromises the CLI that invoked it.

• Resource ceilings behaving as documented. File-size, parse-timeout, candidate and memory limits exist so that a hostile tree cannot run the scan out of resources; a large input that is rejected or truncated with a report entry is working as intended. A ceiling that can be bypassed is in scope.
• Vulnerabilities in the code being analysed. codehelion reports duplication; it is not a vulnerability scanner and does not claim to be.
• Findings that require an attacker to already control the machine the scan runs on, or to have write access to codehelion's own configuration and database.
• Semantic mode running a build script after --allow-execution=build-script was passed. That flag exists to grant exactly that.

First-party code forbids unsafe. Memory-unsafety in a dependency is still in scope for a report — bundled SQLite and the binary-format parsers are C or contain unsafe Rust — and dependency advisories are checked by cargo-deny in CI.

codehelion has not been independently audited.