DeepSeek model nears OpenAI's restricted model in cyber test
In a Vals AI benchmark simulating attacks on two corporate networks, a free open-weight model scored close to a limited-access OpenAI system.
By Standard Topic. AI-generated from the reporting listed under Coverage. Published . Updated .
This story is no longer being updated. It reflects reporting up to .
What happened
Benchmarking firm Vals AI tested AI agents attacking two simulated corporate networks. OpenAI's GPT-6 Sol Daybreak completed about 80.8% of attack objectives, while DeepSeek V4.1 Flash, an open-weight model, reached 72.8%. The DeepSeek model delivered that result at about a tenth of the cost.
Why it matters
Analysts say open models are closing in on closed ones in cyber capability, and open models can no longer be restricted.
Timeline
- Nov 7, 9:38 AM ET: Vals AI results showed OpenAI's GPT-6 Sol completing about 81% of objectives in the simulated network attacks.
- Nov 8, 1:12 AM ET: Vals AI's CyberOffense leaderboard figures put the DeepSeek model's cost at about $3 per run.
Coverage
- DeepSeek Nears OpenAI's Cyber Model in Vals Test at $3 a Run (fourweekmba.com)
More from Technology & AI
- China puts humanoid robots to work serving food and sorting parcels
- Newsom proposes sweeping AI safety rules for Washington
- Google releases Gemini 4 Neon AI model with million-token context
- Microsoft and Intel widen their cloud chip partnership
- Google DeepMind rolls out Genie 4 to select subscribers
- AI helped AMD shorten chip design work, OpenAI says