Prompt Injection Checker
Test your AI agent's exposure to prompt injection before an attacker does.
Paste a prompt, file, or tool output into your AgentGuards dashboard and we scan it against known injection patterns — showing you exactly what an agent would be exposed to.
What we check
- Instruction override attempts — text designed to make the model ignore its system prompt or prior instructions ("ignore previous instructions", role-reassignment, fake system/admin tags)
- Indirect injection vectors — payloads hidden in content the agent will read but the user didn't write: READMEs, tool outputs, web pages, file contents, MCP tool responses
- Encoding/obfuscation tricks — base64, unicode homoglyphs, zero-width characters, and other encodings used to slip instructions past naive keyword filters
- Data exfiltration patterns — instructions that try to get the agent to leak secrets, env vars, credentials, or prior conversation context to an external destination (URL, webhook, DNS lookup)
- Tool-call hijacking — payloads that try to redirect the agent's next tool call (wrong file path, wrong API endpoint, unauthorized command) rather than its text output
- Jailbreak framing — roleplay, hypothetical, or "developer mode" framings used to bypass safety or scope constraints
Why it matters
Coding agents don't just read your prompt — they read everything in their context: files, READMEs, issue comments, web search results, MCP tool outputs. Any of that is an attack surface. A malicious string in a README.md or a poisoned tool response can silently redirect what your agent does next, with your credentials and your repo access.
Most teams test the model. Almost none test what happens when untrusted content the model ingests contains instructions of its own. That's exactly what prompt injection exploits — and it doesn't require compromising your infra, just getting text into somewhere your agent will read it.
How it works
- 1. Paste a prompt, file, or tool output into your dashboard.
- 2. We run it against the pattern categories above.
- 3. You get a breakdown of what was flagged, what category it falls under, and why it's risky.
Want this running automatically?
This page runs a one-off check. Logged-in users get the same scanning running continuously and inline, directly on their agent's live traffic — no copy-pasting required.
Go to your dashboard →AgentGuards runs inline guardrails for AI coding agents — blocking prompt injection, jailbreaks, PII leaks, and data exfiltration via MCP, integrated with Claude Code, GitHub Copilot, OpenAI Codex CLI, and Gemini CLI.