Confident AI — agentic threat model

Uses LLMs (via DeepEval) as evaluators ('LLM-as-a-judge'). Threats include adversarial manipulation of evaluation prompts, bias in the evaluation models, and prompt injection designed to bypass guardrail models.

Manages evaluation datasets, 'golden' test cases, and historical tracing data. Threats include dataset poisoning (to artificially inflate model performance metrics) and the exfiltration of sensitive production data captured in LLM traces.

Orchestrates unit testing, prompt optimization, and guardrail execution. Threats include insecure integration with target LLM applications, manipulation of test execution logic, and evasion of runtime guardrail checks.

Not certain from the listing — likely deployed as a SaaS platform or self-hosted open-source (DeepEval). Threats include unauthorized access to the monitoring dashboard, exposure of API keys used for tracing, and lack of isolation in test execution environments.

This is the core layer of the platform. Threats include evaluation gaming (optimizing prompts to pass specific metrics while remaining unsafe), blind spots in custom guardrail definitions, and drift in evaluation metric accuracy over time.

Not certain from the listing — while it helps other applications achieve compliance, its own internal access controls, RBAC, and data privacy mechanisms (such as scrubbing PII from traces) are not detailed.

Not certain from the listing — primarily acts as an external observer/guardrail rather than an active participant in a multi-agent ecosystem. Threats include cascading latency or denial-of-service in downstream agents if the guardrail/monitoring API experiences outages.