prompt-guard

Prompt injection detection works by scanning untrusted inputs from GitHub issues, emails, and Slack for direct overrides, jailbreaks, and indirect injection. It assigns a severity score and rate limits requests to decide whether to allow, warn, block, or sanitize the message before your agent acts.

Indirect prompt injection is message-based manipulation embedded in external content like emails or GitHub issues. It causes AI agents to follow hostile instructions or reveal credentials by exploiting the agent's context window rather than direct user commands.

You can scan multilingual chat messages for jailbreak attempts and social engineering across English, Korean, Japanese, and Chinese. The system identifies these threats and applies normalization and encoding detection to prevent evasion techniques.

To redact secrets from LLM outputs, you apply output protection and hygiene rules that sanitize model responses before delivery. This process removes sensitive paths and credentials, supporting audit tooling to track what data was blocked.

Prompt injection detection works with GitHub issues, emails, Slack, and Discord messages. It analyzes these agent-facing external sources for untrusted content, applying severity scoring and rate limiting to block malicious instructions before execution.

Severity scoring for AI message security relies on normalization and encoding detection to flag risks, but highly novel or obfuscated attacks may bypass initial filters. Rate limiting and optional threat-intelligence integration help mitigate these edge cases.