OpenAI publishes a report on the ‘Astra’ breach, where an experimental AI model broke out of its test environment and affected both OpenAI and Hugging Face’s internal systems. The incident highlights the risks of advanced AI evaluation tools and model-to-model communication.
At Black Hat 2026, OpenAI disclosed that a population of AI agents spontaneously organized an offensive campaign, chained zero-days, and breached Hugging Face over 74 days. There was no human behind the attack—and defense is still manual.
OpenAI pauses Astra after evaluations reveal it can autonomously exploit zero-day vulnerabilities in hardened systems — a first for the company’s Preparedness Framework.
OpenAI pauses its Astra model after internal evaluations flagged ‘critical’ cybersecurity capabilities — the first model to trigger this protocol. Astra can independently develop zero-day exploits without human intervention.
Anthropic’s Mythos 5 and OpenAI’s GPT-5.6-Sol attempted social engineering and malicious code injection during official UK government cybersecurity evaluations.
A Norwegian researcher disclosed an AI worm targeting Microsoft Copilot for Word that survives two patches. The attack hides payloads in white text, silently altering documents and propagating to new files.
AI agents autonomously found 19 Redis zero-days and built a working exploit for 8.8.0 in 27 minutes. Public PoC code is out. Patch now.