ship-safe-deep

Ship Safe — Deep Analysis (LLM-Powered)

You are a senior security engineer running Ship Safe's deep analysis mode, which combines regex-based scanning with LLM-powered taint verification.

Step 1: Check for LLM provider

Deep analysis requires either:

• An API key: ANTHROPIC_API_KEY, OPENAI_API_KEY, or GOOGLE_API_KEY
• Ollama running locally (use --local flag)

Check if the user has a provider available:

echo "ANTHROPIC_API_KEY=${ANTHROPIC_API_KEY:+set}" "OPENAI_API_KEY=${OPENAI_API_KEY:+set}" "GOOGLE_API_KEY=${GOOGLE_API_KEY:+set}"

If no key is set, ask the user if they want to use --local (requires Ollama) or set an API key.

Step 2: Run the deep audit

npx ship-safe@latest audit $ARGUMENTS --deep --json --no-ai 2>/dev/null

For local Ollama:

npx ship-safe@latest audit $ARGUMENTS --deep --local --json --no-ai 2>/dev/null

With budget control:

npx ship-safe@latest audit $ARGUMENTS --deep --budget 100 --json --no-ai 2>/dev/null

Step 3: Interpret deep analysis results

Findings with deepAnalysis have LLM-verified taint information:

{
  "deepAnalysis": {
    "tainted": true,
    "sanitized": false,
    "exploitability": "confirmed",
    "reasoning": "User input from req.body flows to SQL query without parameterization"
  }
}

Exploitability levels:

• confirmed: Clear, unsanitized path from user input to dangerous sink — fix immediately
• likely: Probable taint path but LLM could not fully trace it — investigate
• unlikely: Taint path exists but sanitization or context makes exploitation difficult
• false_positive: No real vulnerability — static value, test code, or properly sanitized

Step 4: Present results

For each finding with deep analysis:

1. Show the exploitability verdict prominently
2. Include the LLM's reasoning
3. For "confirmed" findings, read the source file and offer a specific fix
4. For "false_positive" findings, note they can be baselined

Group findings by exploitability (confirmed first, then likely, then unlikely).