Sample findings - AgentGuardian

The sibling repo ships pre-captured scan output in sample-findings/ — one directory per agent, with fast-scan.json, fast-scan.sarif, and fast-scan.pdf. Read these before running your own scan to know what the result should look like.

What to look for

fast-scan.json — machine-readable. Top-level findings: [...] array; each finding has category, severity, aivss_score, evidence.prompt, evidence.response.
fast-scan.sarif — GitHub Code Scanning format. Upload to a repo’s Security tab with actions/upload-sarif@v3 and the findings render inline on the PR.
fast-scan.pdf — human-readable report with attack transcripts. The shareable artifact for a launch demo.

Regenerating

The sample files commit empty until the first real scan is captured. To regenerate:

docker compose up --build -d

for agent in vulnerable-langgraph-chatbot vulnerable-mcp-server vulnerable-tool-agent vulnerable-rag-agent defended-baseline; do
  agent-guardian scan \
    --endpoint "http://localhost:8000/${agent}/chat" \
    --model stub --mode fast \
    --output pdf --output-path "sample-findings/${agent}/fast-scan.pdf"
done

Expected severities

Agent	High-severity findings expected	Notes
`vulnerable-langgraph-chatbot`	3-5	LLM01 + LLM06 + LLM07 always; others mode-dependent
`vulnerable-rag-agent`	1-2	LLM02 reliable; LLM10 surfaces in `--mode full`
`vulnerable-tool-agent`	2-3	LLM06 + LLM05 always
`vulnerable-mcp-server`	2-3	LLM01 + LLM06 + LLM07
`defended-baseline`	0	A high-severity finding here is a scanner bug

Use the defended-baseline row as your regression check whenever the scanner ships a new probe library.