If you discover a security vulnerability in prompt-shield, please report it responsibly.

Do NOT open a public GitHub issue for security vulnerabilities.

Instead, please email security concerns to the project maintainers via GitHub's private vulnerability reporting feature on this repository.

• Description of the vulnerability
• Steps to reproduce
• Potential impact
• Suggested fix (if any)

• Acknowledgment: Within 48 hours
• Initial assessment: Within 1 week
• Fix timeline: Depends on severity, typically within 2 weeks for critical issues

prompt-shield is a detection tool, not a guarantee of security. It should be used as one layer in a defense-in-depth strategy.

• By default, prompt-shield never stores raw input text in its database
• Only SHA-256 hashes and vector embeddings are persisted
• The history.store_input_text config option exists but defaults to false
• Threat feed exports contain only hashes and embeddings, never raw attack text

• All data is stored locally (SQLite + ChromaDB files)
• No data is sent to external servers unless you explicitly enable threat feed sync
• The community threat feed is opt-in and pull-only

• The default embedding model (all-MiniLM-L6-v2) runs locally on CPU
• No API calls are made for embedding generation
• Model files are cached locally after first download

We follow responsible disclosure practices. Security researchers who report vulnerabilities responsibly will be credited in the changelog (with permission).