Secure-Code Agent

name: secure-code-agent description: Use when the agent needs to run or respond to secure-code-agent audits, generate bounded security remediation prompts, initialize per-agent security standards files, interpret findings against CWE/OWASP/ASVS/NIST-SSDF, or make small security fixes without crypto-roulette / auth rewrites / validation softening.

Purpose

Use the secure-code-agent CLI as the source of truth for deterministic security audits and bounded remediation prompts. The tool wraps best-in-class scanners (Bandit, Semgrep, pip-audit, npm audit, Gitleaks, built-in regex rules, SARIF ingest from CodeQL/Snyk/Trivy) and normalizes their output across CWE / OWASP Top 10 / OWASP ASVS / NIST SSDF.

The remediation prompt is the differentiator. It encodes hard constraints against the documented anti-patterns for AI security fixes: crypto roulette, auth-flow rewrites, validation softening, test deletion, lint disable, scope creep, dependency thrash, silent behavior change.

Core Workflow

1. Inspect the repo's existing instructions and config before running: - AGENTS.md, CLAUDE.md, .github/copilot-instructions.md, or other agent instruction files. - secure-code-agent.json when present. - .scignore.yaml (suppressions).
2. Prefer the repo's configured command when it exists. Otherwise:

secure-code-agent \
  --config secure-code-agent.json \
  --fail-on-gate \
  --output secure-code-report.md \
  --json-output secure-code-report.json \
  --sarif-output secure-code.sarif \
  --prompt-output secure-code-remediation-prompt.md \
  --comment-output secure-code-pr-comment.md

3. For branch or PR work, pass --changed-only REF (e.g. origin/main). It scans the whole tree and scopes the report to what changed. It issues no grade, deliberately — a scoped run has no denominator it can defend. Pair it with --fail-on-new to gate on regressions.

4. If the audit emits secure-code-remediation-prompt.md, read it before editing and treat it as the bounded task.

5. Fix only the reported findings. Follow the hard constraints verbatim — they're not suggestions.
6. Re-run the audit and any native tests/lints required by the repo.
7. Report the commands run, whether the gate passed, and any remaining false positives or suppression candidates.

Hard rules — non-negotiable

When fixing a security finding:

1. Fix only the findings listed. Do not touch unrelated code.
2. Do not change cryptographic algorithms, key derivation, IV/nonce handling, padding modes, or random sources unless a finding explicitly names them.
3. Do not change authentication flows, session handling, token lifetime, cookie attributes, or authorization gates unless a finding explicitly names them.
4. Do not weaken input validation, output encoding, sanitization, bounds checks, regex strictness, or rate limits to make existing tests pass.
5. Do not disable, delete, or skip security tests. Do not remove @_limiter.limit, @require_auth, @require_csrf, or similar decorators.
6. Do not silence linter warnings via # nosec, # noqa, # type: ignore, eslint-disable, sonar-disable.
7. Do not introduce new third-party dependencies. Prefer stdlib or already-vendored libraries.
8. Preserve behavior. Same inputs → same outputs unless a finding proves current behavior is unsafe.
9. Add a focused test that exercises the security boundary you fixed. The test must FAIL on pre-fix code and PASS on post-fix.
10. Keep the patch small. If you're rewriting a function rather than patching, stop and report.

Installing or running

If secure-code-agent is missing from PATH, check the project docs first:

python3 -m pip install secure-code-agent
python3 -m pip install -e .
python3 -m secure_code_audit.cli --config secure-code-agent.json

Generating per-agent standards

When asked to add AI-agent security standards, use the CLI rather than hand-writing each file:

secure-code-agent --init-agent-standards \
  --target codex --target claude-code --target cursor \
  --target copilot --target windsurf

Generated standards are additive to repository-specific rules. Repo rules win on conflict.

Standards anchors

Every finding carries: - CWE id — the dedupe key (https://cwe.mitre.org/) - OWASP Top 10 bucket (https://owasp.org/Top10/2021/) - OWASP ASVS section (https://github.com/OWASP/ASVS) - NIST SSDF practice (https://csrc.nist.gov/pubs/sp/800/218/final)

Read the linked standard before editing — it is the authoritative description of the weakness, not the scanner's terse rule description.

False positives

If a finding is a false positive, do NOT silently apply a fix. Add a suppression entry to .scignore.yaml with:

• rule_id (or * scoped to file/paths)
• reason (required, non-empty, what's actually safe about this match)
• expires (required, ISO date, ≤ 365 days from today)

The operator will review the suppression in PR.

For interpretation guidance on categories + severity weighting, read references/finding-taxonomy.md.