An open AI model almost matches a top closed one at hacking

Vals, a company that tests the skills of AI models, tried them on test company networks. The Daybreak model of OpenAI, a cyber model for partners that OpenAI trusts, finished 81 percent of tasks. The open model DeepSeek, which a person can download, finished 73 percent. Vals says that at this time open models can do almost the same cyber work as closed models.

How it works

Vals tested the models in two test set-ups. In the first, a model attacked a company network through a website that a person can use. In the second, a model attacked one network to get access to a second network.

The numbers

• Daybreak finished 81 percent of tasks, as an average of two tests.
• DeepSeek finished 73 percent of tasks.
• An attack costs $2 to $3 in token fees with DeepSeek, and $17 to $46 with GPT-6-Sol.

Vals says open models cost less because users can operate them on their hardware. On some tasks the two top models had the same score.

What is not known

The tests used test networks only. In April, Anthropic released Claude Mythos Preview to a small group. It used limits on access to get time. Vals says that these limits are possibly of no use.

Insight: Low cost pulls defenders to open models that teams must check

Vals says low cost gives defenders a strong cause to select open models. In October 2026, researchers made a backdoor in an open model for less than $50. This seems to be a trap: a low cost pulls teams to open models, and each team must then check the source. Leaders who select open models for defense must do that check.