Your agent is still doing its job. Just not the one you built it for.

An agent built to do one thing starts doing something different - not at the level of its general purpose, but at the level of the specific tasks and prompts it executes. The overall goal looks unchanged. The individual actions don’t. That’s where unpredictability lives.

Drift is gradual - the agent expands its behavior over time through workflow changes, new integrations, or ambiguous instructions. Prompt injection is deliberate - malicious instructions embedded in external content redirect the agent to act on an attacker’s behalf. aizome detects and blocks both.

aizome builds a behavioral baseline for every user-agent pair from the moment of onboarding - tracking prompt patterns, tool calls, data accessed, and system interactions. Deviation detection is specific to that agent and that user context, not generic across all agents.

No. Baselines are established automatically. Monitoring runs continuously from the first interaction. No manual threshold-setting or per-agent configuration required.

Onyx and Noma flag intent anomalies. aizome enforces intent boundaries. Detection tells you something went wrong. Enforcement stops it before it executes. That distinction - detection vs. pre-execution enforcement - is the line.

The anomaly is surfaced in the aizome dashboard, the specific action is flagged with context - prompt, tool called, deviation from baseline - and depending on policy, the action is blocked, routed for human review, or logged for audit. The security team gets a complete picture, not just an alert.