AARENA — agentic threat model

The platform's core function is routing prompts to various third-party foundation models. It is highly vulnerable to adversarial prompt injection designed to bypass model guardrails or generate misaligned outputs during public battles.

AARENA collects and stores prompt-response pairs and user voting data. Threats include data poisoning of the evaluation dataset to artificially inflate or deflate specific model rankings, and potential exfiltration of proprietary test prompts.

Not certain from the listing — the specific orchestration framework used to manage model sessions and route anonymous prompts is undisclosed. Potential threats include insecure session handling or prompt leakage between concurrent model battles.

Not certain from the listing — the hosting infrastructure and sandboxing mechanisms for the interactive testing environment are not specified. Threats include API key exposure for integrated LLMs and potential SSRF via model routing endpoints.

As an evaluation platform, its primary threat is evaluation gaming, such as Sybil attacks or coordinated voting campaigns to manipulate leaderboard metrics, alongside blind spots in detecting automated bot interactions.

Not certain from the listing — compliance standards, user authentication, and access controls for premium features are not detailed. Threats include unauthorized access to proprietary evaluation data and lack of audit trails for model interactions.

The platform facilitates a multi-model competitive ecosystem. Threats include cascading failures if upstream model APIs experience outages, and potential trust abuse if a compromised model attempts to exploit the platform's benchmarking interface.