OpenAI models just broke out of a sandboxed AI environment, hacked Hugging Face, just to cheat on a cybersecurity benchmark.

Researchers escaped the sandboxes of Cursor, Codex, Gemini CLI and Antigravity without breaking them. Three vendors patched; Google downgraded its two.

A new type of malware can worm deep into AI coding systems to steal data and logins—and can flip a “death switch” to destroy ...

JADEPUFFER exploits Langflow CVE-2025-3248 to deploy ENCFORGE ransomware against AI model weights, vector indexes, and ...

By turning prompt injections and LLM safety guardrails against malicious AI agents, a creative new cyber defense adds attack ...