DeepSeek Nearly Matches OpenAI Cyber Model in Simulated Network Tests
Reported by Wired and Mezha2 sources ↓
Vals tested AI models on simulated corporate networks and found DeepSeek close to OpenAI’s restricted Daybreak model in completing attack objectives. The results concern mock environments, not real-world intrusions.
Key points
- OpenAI’s Daybreak completed 81% of assigned tasks on average across two tests
- DeepSeek completed 73% of objectives on average
- Vals estimated token costs of $2–$3 for DeepSeek and $17–$46 for GPT-6-Sol
- Daybreak is available only to trusted partners; DeepSeek is open-weight
Sources
Details
Wired reports that Vals tested models in two simulated corporate-network scenarios. One involved a network accessible through a public-facing website; the other required gaining access to one company network through another.
OpenAI’s Daybreak, available only to trusted partners, completed 81 percent of assigned tasks on average across two tests. DeepSeek averaged 73 percent of objectives, and Vals said the models scored equally on some tasks.
Vals estimated that a task cost $17 to $46 in tokens using GPT-6-Sol, compared with $2 to $3 using DeepSeek. The tests measured performance in mock environments; they do not establish how the models perform in real-world intrusions.
Written from coverage as of