Reference
Glossary
Core terms such as agent, session, tool call, trajectory, fail-closed and tenant are defined in Core concepts.
| Term | Definition |
|---|---|
| Indirect prompt injection | Hidden instructions embedded in content the agent reads, meant to steer the agent. |
| Excessive agency | An agent holding more authority than its task needs (OWASP LLM). |
| Egress | Data leaving the trust boundary: tool arguments. |
| Ingress | Data entering the trust boundary: tool results. |
| DLP | Data loss prevention. |
| Drip exfiltration | Slowly leaking data by splitting it across many small calls. |
| MCP | Model Context Protocol; how agents connect to tools and data sources. |
| Shadow mode | An observe-only mode that records decisions without enforcing them. |
| Hash chain | A structure in which each record contains the digest of the previous one, making change detectable. |
| Control plane / Data plane | The management plane / the plane where traffic is actually processed. |