model Bench AI — agentic threat model

The platform connects to over 180 external language models. This exposes it to adversarial prompt injections during evaluation, model output manipulation, and potential model-stealing attacks if users systematically probe proprietary models through the benchmarking interface.

Handles evaluation datasets, prompts, and test suites. Threats include the poisoning of evaluation datasets to artificially inflate or deflate specific model scores, and the exfiltration of proprietary prompts and test cases.

Not certain from the listing — the orchestration framework for managing the 180+ models and prompt optimization tools is not detailed, but insecure integration of model APIs or prompt generation utilities could lead to prompt injection or remote code execution if outputs are unsafely handled.

Not certain from the listing — the hosting environment for the no-code platform and how API keys for the 180+ models are securely stored and sandboxed is unspecified, posing risks of credential theft or container compromise.

This is the core layer of the platform, featuring human and LLM evaluations and output traceability. Threats include evaluation gaming (manipulating LLM-as-a-judge metrics), blind spots in traceability logs, or biased evaluation metrics.

Not certain from the listing — there is no explicit mention of role-based access control (RBAC), audit logging for model evaluations, or compliance with standards like SOC2 or the EU AI Act.

Not certain from the listing — while it connects to 180+ external models, it does not explicitly describe a multi-agent marketplace or collaborative ecosystem, though compromised external model endpoints could feed malicious payloads back into the platform.