Source
Agent provenance
Confirm where the agent came from, who owns it, what package or repo produced it, and which provider and runtime execute it.
Safety
Background checks for AI agents.
Anubis checks where an agent came from, what it can do, how it behaves under pressure, and whether its requested access matches its declared role.
Validation engine
Source, behavior, permissions.
The validation layer focuses on three questions: is the agent's source trustworthy, does it behave safely, and are the requested permissions justified?
Source check
Repository, package signature, dependency risk, runtime, provider, owner, and MCP manifest.
baseline
Prompt injection
Tests whether hostile input can change tool behavior or expose hidden instructions.
high
Data leakage
Checks whether private data can be copied, summarized, or sent outside allowed channels.
critical
Permission fit
Compares requested scopes against task requirements and least-privilege rules.
required
Behavior
Adversarial tests
Run prompt injection, malicious document, tool confusion, and data leakage scenarios before production access.
Access
Permission analysis
Flag standing credentials, broad write scopes, and tools that do not match the declared task.
FAQ
What safety means here.
The purpose is not to promise that an agent is perfect. The purpose is to decide whether the current access request is acceptable, limited, or blocked.
No. Anubis uses validation as part of an access decision. It focuses on source, permissions, tools, data, runtime, and evidence.
Yes. The product is designed around policy enforcement, short-lived access, blocked actions, and revocation when trust changes.
It is a structured review of agent source, owner, provider, runtime, declared purpose, tools, MCP servers, data classes, and requested permissions.