Cyber Security News

Hacking Editorial Brief — August 26, 2026

NVIDIA AI Agent Platform Vulnerable to Drive-By Model Poisoning

A critical vulnerability in NVIDIA's NemoClaw AI agent deployment wrapper (CVE-2026-65105) allows attackers to hijack and persistently poison local AI models through a single malicious website visit. The flaw stems from NemoClaw's default configuration, which binds the underlying Ollama API to 0.0.0.0, exposing it to DNS rebinding attacks. Researchers demonstrated that an attacker-controlled webpage can exploit this misconfiguration to inject malicious instructions, alter model behavior, and maintain persistent access to the AI agent without requiring user interaction beyond visiting the site. The vulnerability highlights emerging attack surfaces as AI agents gain filesystem access and operational autonomy, with security firms warning that autonomous AI systems deployed with overly permissive network configurations create novel exploitation vectors that bypass traditional security boundaries.

Iran-Linked Groups Target U.S. Water Infrastructure and Universities in Coordinated Campaigns

The Department of Justice charged 17 Iranian nationals allegedly affiliated with the Islamic Revolutionary Guard Corps in a multi-year cyber espionage campaign targeting hundreds of U.S. and international universities to exfiltrate research and intellectual property. Separately, Iranian-linked threat groups have been actively exploiting vulnerable programmable logic controllers in attacks against U.S. water treatment facilities, focusing on industrial control systems with known security weaknesses. A separate China-nexus operation dubbed QUICSILVER is targeting Myanmar government personnel and IT sectors using Virtual Hard Disk files disguised as diplomatic event invitations to deliver QUICAgent, a Go-based backdoor that tunnels command-and-control traffic through the QUIC protocol and leverages Cloudflare Workers infrastructure to obscure attacker infrastructure. Additionally, researchers disclosed 24 malicious npm packages abusing unpkg CDN mirrors to redirect developers to ClickFix-style fake CAPTCHA pages for credential theft.

Sources: The Hacker News · Cybersecurity News · Facilities Dive · Campus Reform · Seqrite · The Hacker News