Miscellaneous
Glossary
Terms used across the Axora product documentation.
This glossary explains terms used across the Axora product documentation.
| Term | Meaning in these documents |
|---|---|
| AI agent | Software that uses a model and tools to carry out tasks, potentially including actions on connected systems. |
| AI API gateway | A managed entry point through which applications or agents send model requests and receive responses. Axora Router applies policies at this point. |
| Model channel | A configured model-access path in Router, made available to selected groups. |
| Skill | A component containing instructions or capabilities that an agent can use. Depending on the host, it may include Markdown instructions, scripts, or other resources. |
| Plugin | An extension that adds behavior or capabilities to an agent host. Its integration depth depends on the host. |
| MCP | Model Context Protocol, a protocol through which agent hosts can connect to tools and data sources. |
| RAG | Retrieval-augmented generation: supplying retrieved information, such as knowledge-base content, as model context. |
| Prompt injection | Untrusted instructions that attempt to redirect a model or agent away from the intended task or rules. |
| Runtime protection | Evaluating actions during execution, with interception and enforcement where the integration supports them. |
| Egress | Data leaving a host or controlled environment, including data sent to a model endpoint. |
| Tokenization | In Router's privacy workflow, replacing identified sensitive values with placeholders. This differs from language-model token counting. |
| Content hash | A value used to identify a component's content. AgentGuard uses it with source and version to bind trust to the reviewed component. |
| Fail closed | Withholding permission or a positive decision when the required validation cannot be completed. |
| Air-gapped environment | An environment isolated from external networks. AgentGuard supports private deployment in such environments. |
| Goal | Arena's persisted assessment objective and associated authorization, state, findings, and evidence. |
| Red teaming / adversarial testing | Testing an explicitly authorized scope from an adversarial perspective to identify weaknesses and generate evidence. |
| Observation-only coverage | Visibility into activity without enforceable prevention on that integration path. |
For product-specific boundaries, see Our Product System.