Fiddler AI — agentic threat model

Not certain from the listing — Fiddler uses task-specific, Fiddler-developed trust models. Threats include adversarial evasion of these moderation models or model stealing of Fiddler's proprietary trust models.

Not certain from the listing — Fiddler ingests and processes massive streams of traces, spans, and evaluation data. Threats include data exfiltration of sensitive prompt/response payloads stored in logs, or poisoning of baseline datasets used for drift detection.

Not certain from the listing — Fiddler is an observability tool rather than an agent framework. Threats include insecure integration where client-side orchestration frameworks fail to properly block traffic when Fiddler triggers a guardrail violation.

Fiddler supports cloud and VPC deployments for secure environments. Threats include container compromise or unauthorized network access to the VPC hosting the Fiddler Trust Service, allowing attackers to intercept raw LLM traffic.

This is Fiddler's core layer, offering LLM Guardrails, Traces and Spans, and Fairness tracking. Threats include guardrail bypass (e.g., novel jailbreaks), evaluation gaming, and blind spots in drift detection algorithms.

Fiddler provides centralized controls and responsible governance. Threats include unauthorized administrative access to the Fiddler platform, allowing malicious actors to disable guardrails, alter bias thresholds, or delete audit logs.

Not certain from the listing — Fiddler monitors LLM applications but does not explicitly detail multi-agent ecosystem interactions. Threats include cascading monitoring failures if a compromised upstream agent feeds malformed telemetry that crashes the Fiddler ingestion pipeline.