NVIDIA Eureka — agentic threat model

Relies on OpenAI's GPT-4 foundation model. Vulnerable to prompt injection, adversarial reprogramming, or jailbreaks that could force the agent to generate flawed, malicious, or physically destructive reward algorithms.

Not certain from the listing — likely processes task descriptions, environment states, and simulation logs. Gaps in data lineage or poisoned task descriptions could lead to corrupted reward function generation.

The framework orchestrates an evolutionary loop, autonomously writing and executing Python code (reward functions) in Isaac Gym. This creates a high risk of arbitrary code execution if the LLM output is hijacked or improperly sanitized.

Not certain from the listing — requires integration with NVIDIA Isaac Gym and GPU infrastructure. If the execution environment lacks strict containerization or sandboxing, the dynamically generated code could escape to the host system.

Evaluates reward candidates iteratively based on simulation performance. Vulnerable to reward hacking or evaluation gaming, where the generated code optimizes for simulation metrics while exhibiting dangerous or unintended physical behaviors.

Not certain from the listing — as an open-source research project, it lacks built-in enterprise security controls, access policies, or compliance frameworks, shifting the security burden entirely to the deployer.

Not certain from the listing — primarily operates as a generator-simulator loop. However, the downstream deployment of its trained RL policies onto physical robots represents a critical ecosystem boundary where simulation-to-reality gaps could manifest as physical threats.