Back to radar
Production AI Radar
Prompt injection defenses
Layered controls against indirect and direct injection - especially for agents with tools.
AdoptGovernanceNew
- Why this ring
- Agents with tools are the highest blast radius. Treat injection like XSS: assume untrusted content in every retrieved doc.
- Production risk if ignored
- Injected content triggers tool calls that exfiltrate data or mutate production systems.
- EU AI Act relevance
- Supports cybersecurity and risk management obligations for AI systems with tool access.
- Typical effort
- weeks
- Low FinOps impact
Use cases
- RAG over untrusted docs
- Agent tool calling
- Email/ticket copilots
Adoption steps
- Threat-model retrieval and tool surfaces
- Separate system vs untrusted content channels
- Allowlist tools; human approve writes
- Continuous injection test suite in CI
Related tools
In your assessment
Injection threat model + control effectiveness tests