Threat Intelligence Brief
Curated summary with source attribution
Source: thehackernews.com
Threat Risk: Medium
Victim: Users of NVIDIA NemoClaw and local Ollama deployments
Incident: A vulnerability in NVIDIA NemoClaw allows unauthenticated control of local AI instances via malicious webpages.
Impact: Attackers can plant hidden instructions within AI models to manipulate agent behavior and access tools.
Attacker: Unidentified threat actors
Analysis: The vulnerability arises from insecure network binding of Ollama to 0.0.0.0 on specific platforms, bypassing authentication. Attackers can employ DNS rebinding to trick a user’s browser into sending API requests to the local model server. This allows for the modification of chat templates, enabling the injection of hidden instructions that persist across conversations.
Recommendations: Restrict OLLAMA_HOST to 127.0.0.1 to prevent exposure to the network; Implement strict Host and Origin header validation to mitigate DNS rebinding; Deploy a token-gated reverse proxy for any AI inference backend requiring external access
Source: The Hacker News
Editorial note: this post summarizes third-party reporting and links to the original source.
View Original Source