CLI scanner that treats AI agents, MCP configs, and secrets as suspect

What it does

Ship Safe is a JavaScript CLI that audits repositories for vulnerabilities across application code, CI/CD pipelines, secrets, and supply-chain dependencies. It also targets AI-native risks—prompt injection, agent hijacking, MCP tool poisoning, and RAG context injection—with a fleet of specialized agents that run in parallel and skip irrelevant projects automatically. Core checks run fully offline, while optional AI-backed modes send bounded, masked context to your chosen LLM provider for deeper analysis or red-teaming.

The interesting bit

Instead of bolting AI security onto a generic SAST tool, Ship Safe builds agentic risk into its core architecture with dedicated scanners for things like .cursorrules prompt injection, managed-agent permission drift, and “slopsquatting” (hallucinated package imports). The interactive REPL and fix agent let you review diffs before anything touches disk, and changes remain reversible.

Key highlights

• Covers OWASP categories for LLMs, Agentic AI, CI/CD, and Mobile, plus niche vectors like Roblox asset injection and ML model pickle payloads.
• Runs fully offline for core checks; the --no-ai flag guarantees no data leaves the machine.
• Supports a wide range of LLM providers (Anthropic, OpenAI, DeepSeek, Kimi K3, Ollama, and others) for optional deep analysis and red-team scenarios.
• Interactive REPL with commands for scanning, planning fixes, showing diffs, and undoing changes.
• CI/CD mode outputs SARIF for GitHub code scanning and can gate builds on severity thresholds.

Caveats

• Provider-backed features send context to third-party LLMs after “best-effort credential masking,” so the exact privacy boundary depends on your provider and configuration.
• Many advanced agents target specific ecosystems (e.g., Roblox/Luau, Supabase, Hermes); teams outside those stacks will see those agents skip automatically, but the value proposition varies by codebase.

Verdict

Teams shipping AI-native apps or using MCP servers should look here; it is especially useful if you want local-first scanning with optional AI augmentation. Purely traditional web teams without agentic or LLM surface area might find a more conventional SAST tool sufficient.