Vals AI finds an open-weight DeepSeek model nearly matches OpenAI's restricted cyber model at simulated network attacks

Vals AI · Research report

Read the source (opens in a new tab)
What happened
Vals AI's CyberOffense benchmark page, dated October 23, 2026, reports a test in which seven AI agents attacked two simulated company networks on their own, with no human help.
Key points
  • GPT-6 Sol, which OpenAI offers only through its restricted Daybreak cyber program, scored best overall; open-weight DeepSeek V4.1 Flash 1017 "came close behind."
  • Vals says DeepSeek "scored level with Daybreak on the first network for under $3 per run, against about $22 for Daybreak."
  • On the harder network, GPT-6 Sol Daybreak completed 27 of 31 objectives in under three hours; no model completed that range's final objectives.
  • Vals warns "Once a model's weights are public, no provider can monitor or stop how it is used," and expects such attacks on real enterprise networks to become more common.
Evidence
An independent evaluator's benchmark on two isolated cyber ranges, crediting objectives only from hidden markers found in command output. It measures simulated networks, not attacks on real systems.
Uncertain
Only the top two models' scores on the first network are averages of three runs; the page does not say how many runs the other figures reflect.

See it in the latest record