The flaw was reported to NVIDIA's Product Security Incident Response Team, with details shared ahead of public disclosure. The attack surface is notable because it does not require any prior authentication, meaning a single visit to a crafted page could compromise the integrity of the local AI model.

Key technical points from the disclosure:

• The attack vector relies on the browser or web content interacting with the local Ollama API without proper authorization checks.
• Once exploited, the attacker can issue commands to the Ollama instance and modify the model's behavior persistently.
• The injected instructions remain embedded in the model, potentially affecting all subsequent responses generated by the AI agent.
• Oasis Security emphasizes that this is a supply-chain style risk for developers and users relying on local AI deployments.

No specific CVE identifier was listed in the report, but the issue is understood to affect setups where NemoClaw is deployed with default or permissive network configurations.

Mitigation guidance from the researchers includes restricting network access to the Ollama service, enforcing authentication for local API endpoints, and monitoring model files for unexpected changes.

Source: The Hacker News

Are you running any local AI models exposed to browser traffic, and if so, how are you locking down the API layer against this type of attack?