Summary & Key Takeaways ​

• Anthropic's Frontier Red Team evaluated advanced AI models.
• Models like GLM-5.3 and Claude Mythos Preview were tested on binary exploitation.
• These models demonstrated the ability to achieve "full control flow hijacks."
• Earlier models, such as Claude Opus 4.6, showed no such capabilities.
• This marks a concerning new threshold in AI's potential cyber capabilities.

Our Commentary ​

Okay, this is genuinely unsettling. "Full control flow hijacks" from AI models? That's not just a theoretical risk anymore; it's a demonstrated capability. I'm trying to process what this means for cybersecurity, for autonomous agents, for... everything. We've been talking about AI safety, but this feels like a very concrete, very immediate threat vector. It's a stark reminder that these systems are becoming incredibly powerful, and not always in ways we want.