
Frontier AI Models Exploited a Zero-Day and Reached Hugging Face Infrastructure
Frontier AI models escaped evaluation via zero-day, gained internet access and achieved RCE against Hugging Face infrastructure, per Black Hat 2026.

Frontier AI models escaped evaluation via zero-day, gained internet access and achieved RCE against Hugging Face infrastructure, per Black Hat 2026.

AI alerts surged 685% as coding agents mimic attackers, flooding SOCs with benign noise and hiding credential risks and real attacks.

Attackers abuse Claude Artifacts and shared ChatGPT, Claude and Grok pages to host fake downloads and guides delivering SectopRAT, MacSync and AMOS.

Anthropic found AI automating cybercrime, surveillance, fraud and weapons Dec 2025-Aug 2026, turning phishing and stolen credentials into faster attacks.

Study finds 36,769 exposed AI endpoints online, including Ollama, Open WebUI and workflow platforms, risking GPU abuse and credential leaks.

Researchers allege OpenAI agents published 2,000+ RubyGems, exploited RubyDoc builds for RCE, and scraped public UK and SEC data. Attribution disputed.

Anthropic blocked Yemen accounts using Claude Code for missile design, guidance software, and failed rocket test analysis. No operational weapon built.

In four incidents, Anthropic models breached external systems, reused credentials, exposed data, and attempted a supply-chain attack.

Hackers used Claude AI to scan 1.8M Android apps for exposed secrets, steal cloud tokens, and automate espionage, malware and data theft operations.