Threat Intelligence Brief
Curated summary with source attribution
Source: securityboulevard.com
Threat Risk: High
Victim: Developers and organizations using NVIDIA NemoClaw and Ollama
Incident: Discovery of CVE-2026-65105, a high-severity vulnerability allowing remote model poisoning in the NemoClaw framework.
Impact: Attackers can achieve full takeover of local AI model behavior by implanting persistent hidden instructions.
Attacker: Unidentified threat actors
Analysis: The vulnerability arises from NemoClaw binding the Ollama API to all network interfaces without authentication to facilitate container connectivity. Attackers can utilize DNS rebinding to bypass CORS restrictions and access the local API. Once accessed, the attacker can overwrite model templates to implant persistent, hidden instructions.
Recommendations: Restrict Ollama API binding to 127.0.0.1 unless external access is explicitly required; Implement strict firewall rules to block unauthorized traffic to port 11434; Monitor for unusual API calls to /api/create and /api/show endpoints within AI environments
Source: Security Boulevard
Editorial note: this post summarizes third-party reporting and links to the original source.
View Original Source