Threat-model and remediate code with ak:security

Use ak:security for a secret, dependency, and code-pattern scan plus STRIDE and OWASP threat modeling. Default invocation is that comprehensive pass. --secrets-only and --deps-only skip code patterns and threat modeling; combining them still skips code patterns. Optional red-team and fix modes add deeper discovery and local code changes.

Choose ak:security for threat-modelled review

Use ak:security when

• Authentication, authorization, payments, personal data, or sensitive system boundaries have changed.
• You want findings connected to attacker scenarios and source locations.
• You need external, supply-chain, insider, and infrastructure perspectives.
• You are preparing for a release or a focused compliance review.

Choose another workflow when

• You only need a fast secret or dependency pass. Stay on ak:security and pass --secrets-only, --deps-only, or both.
• The change is cosmetic and does not affect data or user-facing behavior.
• You need a real penetration test against a running service; this Skill is a source-led audit and does not claim exploit execution.
• You already have a confirmed defect and only need implementation. Use ak:fix with the evidence and acceptance boundary.

Prepare a bounded audit surface

Before you start:

• Complete Onboarding, and confirm Engineer Kit is installed for the current runtime and scope.
• Provide a file glob, directory, or the exact keyword full.
• Make the relevant source, manifests, lockfiles, CI configuration, deployment files, and repository instructions available.
• Install the stack's audit tool if you expect dependency evidence: npm audit, pip-audit, govulncheck, or bundle audit for the documented stacks.
• Decide whether network-backed advisory lookups and local fix commits are allowed before selecting their modes.

Select depth and mutation separately

/ak:security --secrets-only
/ak:security --deps-only
/ak:security src/auth/ --red-team --iterations 12

/ak:security --secrets-only
/ak:security --deps-only
/ak:security src/auth/ --red-team --iterations 12

$ak:security --secrets-only
$ak:security --deps-only
$ak:security src/auth/ --red-team --iterations 12

Start without --fix when you need to calibrate findings or obtain approval for risky remediations. A realistic request also names trusted test commands, files that must not change, and whether local commits are acceptable.

Understand the observable stages

1. The Skill resolves scope. It expands the glob, directory, or full and reads every selected file before analysis.
2. It maps threats. STRIDE covers spoofing, tampering, repudiation, information disclosure, denial of service, and elevation of privilege.
3. It checks OWASP categories. Access control, cryptography, injection, design, configuration, components, authentication, integrity, logging, and SSRF receive an explicit pass.
4. It audits dependencies and secrets. The detected stack selects a package auditor, while source is checked for credential patterns.
5. Red-team mode iterates personas. External adversary, supply-chain, insider, and infrastructure phases select vectors, trace flows, validate evidence, and chain relevant findings.
6. It consolidates evidence. Findings are deduplicated, assigned Critical, High, Medium, Low, or Info severity, and written into the report.
7. Fix mode remediates iteratively. Each targeted change runs a test or lint guard and creates a security(fix-N): ... commit only after the guard passes.

Optional finding-triage experiment

Ordinary review is unchanged and makes no provider call. Only when you explicitly opt into a finding-triage experiment after candidate discovery does the Skill use ak eval triage, whose default mode is off. Shadow mode requires your provider consent, explicit review-triage consumer consent, and a credential. The Skill sends only compact, sanitized facts: it removes every discovered secret value from the batch and never submits raw scanner output, full source, credentials, or sensitive paths.

The shadow signals are compared with the original decisions and never replace them. Protected Critical or High signals, unknowns, and provider failures keep deep review, and duplicate or auto-fix signals never authorize suppressing a finding or editing code. To measure the experiment offline, use ak eval triage-benchmark with frozen shadow receipts and independently verified labels.

Protect credentials and repository state

• Dependency audit commands can contact advisory registries and update local caches. Their availability, data transfer, rate limits, and provider cost are properties of the selected tool and environment.
• Static inspection does not authorize requests against production systems, credential use, exploit execution, data exfiltration, or destructive testing.
• --fix authorizes local edits and one commit per successful fix iteration. It does not authorize push, pull request creation, merge, release, or deployment.
• A failing guard stops the fix loop and remains visible; the Skill does not skip or weaken verification to continue.
• Review generated changes before accepting broad security hardening. Some findings require a product, compliance, or compatibility decision that automation cannot make.

Verify the report and fixes

A complete audit should provide:

• Number of files scanned and finding counts by severity.
• Category, source location, description, attack scenario, impact, confidence, and a bounded fix recommendation for each finding.
• Red-team persona and iteration evidence when that mode runs.
• Dependency-audit command and result, or an explicit unavailable reason.
• Redacted secret evidence as an opaque marker, never a partial live secret.
• For fix mode, changed files, guard commands and results, commits created, stopped iterations, and unresolved findings.

The packaged evidence confirms installation coverage for the Skill; it does not provide an end-to-end test of the full audit, live advisory services, or fix quality.

Troubleshoot and interpret limits

The retired ak:security-scan invocation maps to ak:security. Continue with ak:plan to schedule findings that should not be fixed immediately.