Skip to main content

AI-code audit

By 2027, 40-60% of code submissions will be AI-generated. The regulated-industry compliance question - "which findings sit on AI-touched code?" - needs a real answer. Vulkro is the only SAST tool that ships a defender-side AI-origin audit today.

This page covers the three CLI surfaces that consume the audit:

  • AI-origin fingerprinting (per-tool markers: Claude, Copilot, Cursor, Aider, ChatGPT, generic AI-generated)
  • --ai-code-segregation segregation report
  • --attest-reviewed reviewer attestation

Why this is on the SAST tool​

The traditional security-tool answer is "we run the same detectors on every line". That's correct for technique but wrong for posture: AI-generated code is empirically more prone to specific failure modes - boilerplate auth that does not enforce, copy-paste secrets, missing error handling, hallucinated dependency names. The right place to surface "this finding is on AI-touched code" is inside the scanner that already knows about the finding.

Vulkro stays on the defender side: this module audits AI-generated code; it does not generate fixes via a vendor LLM. That matches the offline-first posture - your code never leaves the host.

How origin is detected​

src/security/ai_origin.rs scans each module for canonical per-tool markers:

ToolMarkers it looks for
ClaudeGenerated by Claude, Generated with Claude, Co-Authored-By: Claude, @claude-generated, @anthropic-generated
GitHub CopilotGitHub Copilot, Copilot-generated, @copilot
Cursorcursor.sh, @cursor-generated
Aideraider:, // aider
ChatGPTGenerated by ChatGPT, OpenAI GPT-...
GenericAI-generated, Generated by AI, # generated by ..., // generated by ...

The marker scan is strict precision today - if the source doesn't carry an explicit marker, the file is treated as hand-written. Heuristic shape detection (long boilerplate docstrings, uniform variable naming) is tracked for a follow-up; the strict marker scan is the safer first step because false-positive "this was written by Copilot" findings on hand-written files would be worse than the current under-coverage.

Marker line numbers are preserved so a per-file marker-ratio ("3 of 280 lines carry an AI marker") can be computed.

--ai-code-segregation​

Adds a markdown report after the scan output:

$ vulkro scan . --ai-code-segregation
... (normal scan output) ...

# Vulkro AI-Code Segregation Report

- Total findings: 42
- Findings on AI-touched files: 11 (26.2%)

## Per-tool breakdown

- claude: 7 finding(s)
- copilot: 3 finding(s)
- generic-ai: 1 finding(s)

## AI-touched files (marker-line ratio)

- src/api/users.ts: 0.71%
- src/api/auth.ts: 1.43%
- src/services/notifier.ts: 0.36%

The report is emitted to stderr; the main output stream (table / JSON / SARIF) is unchanged so machine-readable formats stay parse-clean.

--attest-reviewed​

Stamps every finding on an AI-touched file with a human-reviewed-ai-code evidence row. Lets a downstream compliance emit show "every AI-touched finding has a human-review sign-off" without per-finding manual tagging.

$ vulkro scan . --attest-reviewed --reviewer jane@team
... (normal scan output) ...
✓ attest-reviewed: stamped 11 finding(s) on AI-touched files (reviewer: jane@team)

The reviewer name defaults to $USER; pass --reviewer NAME to override. The attestation lives on evidence[].signal = "human-reviewed-ai-code" and evidence[].detail = "Reviewer attestation: AI-touched code reviewed by ...", so it persists into SARIF, JSON, JUnit, and the database emit.

Use cases​

HIPAA / PCI-DSS / FedRAMP audit evidence​

The auditor asks "show me that every AI-touched line has been reviewed by a human". Run vulkro scan . --ai-code-segregation --attest-reviewed --format sarif > scan.sarif. The SARIF carries both the per-tool breakdown (properties rolling up segregation counts) and the per-finding attestation evidence row. Hand the auditor the SARIF file.

Pre-merge gate on AI-shop PRs​

CI runs vulkro scan . --ai-code-segregation on every PR. The stderr report goes into the PR's CI log; if the AI-touched-share exceeds the team's threshold, a wrapping script fails the build and posts the report as a PR comment.

Forensics after a security incident​

A security event surfaces in production. Run vulkro scan . --ai-code-segregation against the commit that shipped the bug. The output tells you which tool produced the offending code, so the team can update its tool-specific reviewer playbook.

Last updated on