Tool · Security scan
Is that text telling your agent what to do?
Every email, page and document your agent reads is also a set of candidate instructions. This scanner flags the known patterns - instruction overrides, chat-markup impersonation, exfiltration endpoints, secrecy demands, fake authorization - before the text meets the model. Runs entirely in your browser: nothing is sent anywhere.
Paste the text your agent is about to read - an email, a web page, a document, a chat export - or load it as a file. The scan runs in this browser tab and the text is never sent anywhere.
Sources
- OWASP Gen AI Security Project: LLM01 Prompt Injection
- OWASP Top 10 for Large Language Model Applications
- Simon Willison: Prompt injection (series)
How a poisoned argument moves through an agent's tools: Parameter poisoning and the AuthorityFlow defense, and how injected memory survives between sessions: MemHarmBench and cross-session prompt injection. Lint the config that decides which tools the agent can reach with the MCP config lint.