Security-Focused Guide for AI Code Assistant Instructions (best.openssf.org)

The OpenSSF Best Practices and AI/ML Working Groups published a practical guide showing how to harden AI code assistants by feeding them concise, security-focused custom instructions (e.g., Claude markdown, Copilot files, Cursor/Kiro rules). The guide’s point: steer models to treat security and supply-chain hygiene as first-class concerns so generated code is less likely to introduce vulnerabilities. It includes copy-paste-ready directives and a short checklist covering input validation, parameterized queries, escaping/encoding for HTML/SQL, secret-handling via env/vaults, safe auth flows, role checks, constant-time comparisons, safe defaults (HTTPS, strong crypto), and mandatory security reviews for TODOs or placeholder code. Technically, the guidance spans tooling and platform practices: pin dependency versions and prefer official package managers, produce SBOMs (SPDX/CycloneDX), use in-toto or attestation frameworks, verify container images with cosign/notation and immutable digests, enable admission controllers in Kubernetes, and integrate SAST/DAST/dependency scanners (CodeQL, Bandit, Semgrep, OWASP Dependency-Check) in CI. Language-specific rules cover C/C++ buffer safety and compiler hardening, Rust’s avoidance of unsafe blocks, Go race detection, Python avoidance of eval/exec, and safe JS/Java/.NET crypto/identity libraries. The guide stresses concise, actionable prompts (avoid overloading models), iterative review of assistant outputs, and continued experimentation — enabling teams to embed a “security conscience” into their AI-assisted development workflows.