SynthScan

An AI security agent that knows the specific vulnerability fingerprints of each major AI coding tool — Claude Code, Copilot, Cursor, Antigravity — and automatically reviews every PR containing AI-generated code with that context in mind.

Problem

Generic SAST tools (Semgrep, Snyk, Aikido) apply the same rule set to every line of code. But Georgia Tech's Vibe Security Radar has tracked 74 CVEs directly attributable to AI coding tools as of March 2026 — with March 2026 alone producing 35, more than all of 2025 combined. The root cause: AI coding tools generate systematic vulnerabilities. Tenzai's December 2025 study of 15 AI-built apps found 100% introduced SSRF vulnerabilities and 0% implemented CSRF protection. OX Security found 62% of AI-generated code ships with exploitable flaws. Generic pattern-matching scanners catch isolated instances; they don't know that Copilot reliably skips CSRF, that Claude Code leaves specific command-injection signatures, or that 19.7% of AI-generated import statements reference hallucinated package names that don't exist on npm or PyPI — a direct dependency-confusion attack surface.

Target user

Backend or full-stack developer at a startup or mid-sized engineering team who ships 40–60% of their production code via Claude Code, Copilot, or Cursor. They've read the vibe coding breach reports, know they should be doing better security review, but don't have a security engineer on staff and don't want to manually audit every AI-assisted PR. They need something that plugs into GitHub, runs automatically, and surfaces only real findings — not a sea of false positives from a generic linter.

MVP scope

• GitHub Action that detects AI-generated code in PRs via commit trailers (Co-Authored-By: github-actions[bot], claude, copilot), Antigravity/agy metadata headers, and file-change heuristics for common AI scaffolding patterns
• Sends flagged files to a security-specialized LLM with a per-tool vulnerability profile loaded into context (Claude Code: injection patterns and RLS omissions; Copilot: CSRF/auth skips; Cursor: insecure dependency resolution; Antigravity: SSRF in generated server routes)
• Package hallucination detector: extracts all import and require statements, checks each against npm/PyPI/crates.io registries, and blocks the PR merge if a non-existent package is found
• Posts a structured PR review comment: severity, CWE, OWASP category, affected line range, one-sentence explanation, and a suggested fix patch
• PR-level summary block: "SynthScan reviewed 8 of 12 files (64% AI-generated) — 3 findings: 1 critical (SSRF in /api/fetch), 1 high (missing CSRF token on POST /user/update), 1 info (debug logging in production path)"

Monetization

Freemium GitHub App. Free tier: 30 PR reviews per month, basic LLM reviewer (haiku-tier model), community vulnerability profiles for Claude Code and Copilot. Pro at $15/month: unlimited reviews, priority LLM queue (sonnet-tier model), team-configurable suppression rules, Slack and JIRA integrations, monthly AI code security digest for the org. Team at $49/month: org-wide deployment via GitHub policy, audit-ready PDF report per quarter, custom vulnerability profiles for internal AI coding tooling.

Why now

Four signals converged in the first half of 2026. First, Georgia Tech's Vibe Security Radar documented 74 CVEs attributable to AI coding tools by March 2026, with the rate accelerating sharply: 18 cases in all of 2025, 56 in Q1 2026, 35 in March alone. Second, OX Security's June 2026 analysis found 62% of AI-generated code ships with exploitable vulnerabilities — 86% fail XSS defenses, 2.74x more cross-site scripting than human-written code. Third, Google Antigravity (launched May 19, 2026) and GitHub Copilot's switch to usage-based AI Credits billing (June 1, 2026) are pushing more AI-generated code into more repos than ever. Fourth, EU AI Act enforcement begins August 2026, creating organizational pressure to document which code was AI-generated and whether it received a security review before shipping.

Risks & open questions

• AI-tool attribution is noisy: commits from teams that squash or rebase strip the Co-Authored-By trailers; false-negative rate for tool detection could be 30–40% without heuristic fallbacks
• LLM security reviewers have non-trivial false-positive rates without careful calibration — a noisy PR bot gets ignored or uninstalled within a week
• Build-side: inference latency on large PRs (50+ files) could push CI wait times past 3–4 minutes, which is a hard adoption ceiling for many teams
• Generic SAST vendors (Snyk, Semgrep) are adding AI-aware rule sets; if they ship good AI-tool-specific profiles before SynthScan gets traction, the differentiation narrows
• Demand-side: teams with existing SAST pipelines may perceive SynthScan as redundant even when the per-tool vulnerability profiling is genuinely additive

Next step

Build a 200-line proof-of-concept GitHub Action that detects Claude Code commit trailers, sends the diff to a security LLM with a tight system prompt, and posts a comment — test on 10 recent open-source repos with known AI-generated CVEs from the Vibe Security Radar dataset to measure precision.

Sources

• https://www.ox.security/blog/vibe-coding-security/ — 62% of AI-generated code ships with vulnerabilities; 86% XSS failure rate; 2.74x cross-site scripting vs. human code; Escape.tech analysis of 5,600 vibe-coded apps (June 2026)
• https://research.gatech.edu/bad-vibes-ai-generated-code-vulnerable-researchers-warn — Georgia Tech Vibe Security Radar: 74 CVEs tracked as of March 2026, up from 18 in all of 2025; Claude Code and Copilot leave the clearest tool signatures