Every sample on these pages comes from running the scanner on an example input. Nothing is written by hand.
Find out whether an MCP server answers without credentials or grants more power than it should, before your agent connects.
A verdict (clear, caution or high-risk), a 0 to 100 risk score, and one finding per problem with the tool it came from.
scan_mcp_server
Catch hidden instructions in a document or web page before an agent reads it.
A verdict and risk score, with a finding for each hidden or obfuscated payload and each page element built to steer a browsing agent.
scan_content
Check a Claude Skill or a VS Code extension for risky grants and untrusted publishers before you install it.
A verdict and findings. For a skill: shell access it asks for, commands it runs on load, hidden text. For an extension: publisher domain verification weighed against installs.
scan_extension
Flag agent output that parses as valid JSON but holds no real answer.
A verdict (looks-real, suspicious or degenerate) and a finding for each placeholder value, missing field and run of identical items, with its path.
check_output