asamassekou10/ship-safe

Ship Safe is a local security scanner (CLI) built specifically to find and verify vulnerabilities in software, with a heavy focus on AI-written code and AI agent configurations. It runs in your repository, finds potential issues, and then investigates each one to prove whether it is a real threat, showing you the exact lines of code and reasoning behind its conclusion.

What it does: It works in two layers. First, a fast, deterministic engine scans your codebase for issues like leaked secrets, SQL injection, supply-chain risks, and AI-specific threats (prompt injection, agent hijacking, RAG poisoning, MCP misconfigurations). Second, an investigation layer verifies each finding—it traces whether a dangerous value actually reaches a vulnerable function, searches for missing security controls, builds attack chains across multiple config files, and can even probe a leaked API key against its provider to confirm it works. Every finding is labeled confirmed, likely, unresolved, or refuted, with the evidence trail cited.

Key features:

• Evidence-based verdicts: It ranks its own investigation methods (e.g., a traced data path outranks an AI model's guess) so you can trust the conclusion.
• AI/LLM security focus: Detects OWASP LLM Top 10 and Agentic AI Top 10 risks, including prompt injection, excessive agency, memory poisoning, and malicious MCP servers.
• Agent capability mapping: The capabilities command reads your agent configs (like CLAUDE.md and .claude/settings.json) to show what an AI agent could actually reach and exploit.
• Interactive REPL and fix agent: You can scan, ask questions, and let the tool propose a diff to fix a vulnerability—which you approve before it writes.
• CI/CD integration: Fails builds on critical findings, outputs SARIF for GitHub code scanning, and supports gating on evidence (e.g., only block on confirmed verdicts).
• Local and private: Core scanning works offline without an API key; AI-backed analysis uses your own configured LLM provider (Anthropic, OpenAI, Kimi, etc.).

Tech highlights:

• Node.js/npm CLI (MIT licensed).
• Supports JavaScript, TypeScript, Python, config files, and infrastructure-as-code.
• Exposes a local MCP (Model Context Protocol) server so AI coding clients like Codex, Claude Desktop, and Cursor can use it.
• Includes a benchmark suite that scores how well it separates real findings from noise.

Use cases:

• Running a security audit before opening a PR or merging AI-generated code.
• Checking a downloaded or unfamiliar repo for malicious instructions or dangerous agent permissions before opening it with an AI tool.
• Gating CI/CD pipelines to block only verified, high-severity vulnerabilities.
• Securing AI-native applications that use agents, RAG pipelines, or MCP servers.