SECURITY.md

Security Model and Reporting

Security boundaries

The principal attack surfaces are:

• untrusted .emath source and package metadata;
• hostile custom-kind schemas and lowering rules;
• dependency/registry substitution;
• malicious provider plugins;
• solver output and proof/certificate forgery;
• generated Rust and build-script execution;
• JIT/GPU/FFI backends;
• resource exhaustion and adversarial symbolic growth;
• evidence or source-map tampering;
• host promotion based on poisoned measurements.

Default posture

• Parsing and semantic admission are bounded and fail closed.
• Package resolution uses content locks.
• Compiler plugins run in a sandboxed component boundary unless statically trusted.
• Native providers are explicit build dependencies and appear in the trust manifest.
• Build scripts and procedural macros are treated as code execution.
• Generated Rust defaults to #![forbid(unsafe_code)].
• Solvers have proposal authority; independent checkers grant evidence authority.
• All expensive phases consume explicit budgets and cancellation tokens.
• Artifact manifests and evidence bundles are content-addressed.
• Promotion requires frozen workloads, provenance and rollback.

Reporting

A production repository should add a private reporting address and coordinated disclosure policy before public release. Do not open public issues containing exploitable inputs, tokens, private model artifacts or unpublished proof failures.