AI News
All stories
Reported

Vals test: AI agents breach realistic enterprise networks; GPT-6 Sol leads at 80.8%

Vals CyberOffense found AI agents autonomously breached realistic enterprise networks, with GPT-6 Sol completing 80.8% of objectives and DeepSeek V4.1 Flash 72.8%.

Safety & Security · · 1 min read

Watch the short

Under a minute, read by an AI anchor. Made .

YouTube loads only when you press play.

Vals reported that AI agents autonomously breached realistic enterprise networks in its CyberOffense benchmark. GPT-6 Sol completed 80.8% of objectives, and the open-weight DeepSeek V4.1 Flash completed 72.8%. Outside developers cannot download or freely call the top scorer, since GPT-6 Sol is available only through OpenAI's restricted Daybreak program. The 8-point gap between a closed model and an open-weight one is the figure to watch.

In the lighthouse scenario, Daybreak completed 27 of 31 objectives in under three hours, while DeepSeek completed 21 objectives at about a tenth of the cost. Vals put DeepSeek at under $3 per run, against about $22 for the Daybreak run.

Vals also found that DeepSeek made most of its progress after 100M tokens, so short tests understate open models. A test that stops early would capture only part of what the open model can do.

Why it matters

Autonomous intrusion capability is now measured for both restricted and open-weight models. Security teams should assume low-cost open models can attempt similar attacks.

What you can do

Red-team internal networks with agent-based tests, and check which AI tools have network access that an agent could misuse.

Original source

We have no link to the original source for this story yet.

Where we saw it

1 source

Latent Space

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works