Guardian AI Lab
Open projects for practical safeguards.
Ethical foresight and security overlap, but they require different tests.
Prompt-injection defenses
Detect and contain instructions smuggled through untrusted content.
Permission enforcement
Keep actions inside delegated authority, regardless of capability.
Uncertainty monitoring
Flag decisions made on weak or stale information.
Omission detection
Notice when needed beneficial action is not taken.
Independent oversight
Separate the system that acts from the one that reviews.
Rollback
Make consequential actions reversible wherever possible.
Safe self-modification
Safeguards for AI-assisted changes to the systems themselves.
Contribution standards
- Threat models
- Reproducible benchmarks
- Failure reports
- Versioned contributions
- Explicit licenses