Ship Safe is a free, MIT-licensed AI code security agent that acts as an independent reviewer for AI-written code — a deterministic scanner finds issues the coding agent cannot see in its own work. This cybersecurity news check tests your AI code review posture.
— Ingenious AI security check.
Asking the coding agent to review its own work is marking its own homework. Most AI-generated code ships with no independent review at all — the author cannot say what it can reach, what its permissions combine into, or whether it is consistent. The check below grades your review pipeline in 60 seconds.
An automated reviewer that scans AI-generated code for security flaws independently of the agent that wrote it — checking what the code can reach, what its permissions combine into, and whether it is consistent.
Asking the coding agent to review its own work is marking its own homework. Ship Safe is a free, MIT-licensed AI code security agent that acts as an independent reviewer: a deterministic scanner finds issues across your code, agents, MCP configs, and CI — then an investigation layer decides whether each finding is real, tracing the evidence line by line. Five questions on whether your AI-written code is actually reviewed.
Short answer: most AI-generated code ships with no independent review at all — the author (an AI) can't answer what it can reach, what its permissions combine into, or whether it's consistent. Ship Safe answers all three from outside the agent, with evidence you can argue with. Run npx ship-safe — no signup, no API key. Check your five below.
Answer for the codebase you ship.
| Layer | What it does |
|---|---|
| Deterministic engine (sensor) | Fast, repeatable scans of app code, AI agents, MCP configs, prompts, dependencies, CI/CD, secrets, cloud-adjacent config. Same input → same output. |
| Investigation layer (judge) | Traces tainted values to sinks, searches for missing controls, builds attack chains across files no single review sees. Verdicts: confirmed / likely / unresolved / refuted — each citing the lines it read. |
| Fix agent | Proposes plan + diff, asks before writing, reports static rescan, reversible with undo. |
| CI gate | ship-safe ci fails risky builds, uploads SARIF to GitHub code scanning. Gate on severity or on evidence verdicts — e.g. block only confirmed findings. |
npx ship-safe audit . (full audit), investigate . (evidence verdicts), trust ~/Downloads/take-home (what runs when you open an unfamiliar repo with an agent), capabilities . (what an agent in this repo can actually reach).--no-ai guarantees a fully local scan.Sources: asamassekou10/ship-safe (GitHub) · shipsafe.sh
A single-file review can't see that your CLAUDE.md, your agent's launch permissions, and an MCP write tool combine into a path from a pull request to a privileged write. ship-safe capabilities reads all of it from outside.
Severity-only gates drown teams in noise. Gating on --fail-on-verdict confirmed blocks what was established by evidence and lets argued-away findings through.
Compare a trusted base scan against the PR head so existing repository debt stays visible without blocking unrelated changes — the repo documents the exact workflow.
Run a static-analysis (SAST) scanner over the repo — open-source options include Semgrep and vibe-code-focused AI code security scanners — plus a secrets scan for hardcoded API keys, then a dynamic (DAST) scan against the running app. No scanner replaces a human review of authentication, data handling, and business logic.
Vibe coding security is the practice of checking AI-generated code for the flaws models commonly introduce: hardcoded secrets, injection flaws, cross-site scripting, disabled TLS verification, and insecure defaults. Research (Veracode's 2025 GenAI Code Security Report) found 45% of AI-generated code shipped a real vulnerability.
A free, MIT-licensed security agent for AI-written software (npx ship-safe). It combines a deterministic scanning engine — covering application code, AI agents, MCP configs, prompts, dependencies, CI/CD, secrets, and cloud configuration — with an investigation layer that traces data paths and resolves findings to confirmed, likely, unresolved, or refuted, each with cited evidence.
Three structural differences: it's a separate reviewer rather than the author checking its own work; it can see what an agent can reach (MCP server configs, launch permissions, tool grants) from outside; and it's deterministic — the same input produces the same output, so results can be gated in CI and improvement can be proven over time.
Core checks work offline with no signup and no API key. Use --no-ai to guarantee a fully local scan. AI-backed red-team modes use your configured provider when available, with best-effort credential masking.
Yes — ship-safe ci fails risky builds and uploads SARIF to GitHub code scanning, with a GitHub Action for inline PR findings. You can gate on severity or on evidence verdicts (e.g. block only confirmed findings), and compare against a baseline so existing repo debt doesn't block unrelated changes.
Open-source AI code security scanners include vibescan, isitsecure, and AXguard, plus Cloudflare's security-audit-skill and general tools like Semgrep for static analysis and OWASP ZAP for dynamic testing. The right pick depends on your stack: SAST for code, DAST for running apps, secrets scanners for leaked keys.