AI agents breach corporate networks
The story so far
Vals AI finds AI agents can break into realistic corporate networks on their own. OpenAI's GPT-6 Sol Daybreak, part of a restricted project, achieved 80.8% of attack objectives, while the open-weight model Deepseek V4.1 Flash 1017 reached 72.8%. Both models operate autonomously on isolated machines, sending commands from web applications to Windows servers for admin privileges. Vals researchers note that once a model's weights are public, no provider can control or restrict its use, and most models need over 100 million tokens to make meaningful progress, meaning short tests underestimate open models' potential. Deepseek V4.1 Flash 1017 costs about $3 per run, significantly less than the roughly $22 for GPT-6 Sol Daybreak, and scored zero results in its first 25 million tokens, improving only after reaching that threshold.
AI-written from the reports · Updated 11 days ago
Coverage timeline
Follow the coverage to see every side of the story.
- The DecoderThe gap between open and closed AI models in cyber capabilities keeps shrinking
Vals AI's CyberOffense benchmark allows AI agents to autonomously attack simulated enterprise networks on isolated Kali Linux machines, GPT-6 Sol Daybreak in OpenAI restricted projects completes 80.8% of the target, open-source weighted model Deepseek V4.1 Flash 1017 reaches 72.8%, with a single run cost of about 3 dollars instead of about 22 dollars