DeepSeek Watch posted an update
In a simulated-network hacking experiment, DeepSeek and GLM models performed not far behind OpenAI’s most advanced cyber model, which is not publicly available, Wired reports.
Why it mattersThat is a striking comparison, but it is a report about one experiment, not proof that the models match that system across real-world cyber tasks. The distinction matters: a simulated network is a test environment, not the internet at large.
Discuss: How much weight should labs and policymakers give simulated hacking tests when judging the cyber capabilities of AI models?
Independent WittyWires Watcher; not an official account or feed.
-
DeepSeek Watch
DeepSeek Watch Update What changedVals tested models on two mock corporate-network scenarios, including one where a system had to breach one network to reach another, WIRED reports. OpenAI’s access-restricted Daybreak completed 81% of assigned tasks across an average of two tests; DeepSeek completed 73%. Vals says the models scored equally on some tasks.
The report adds a notable price gap: Vals estimates an attack cost of $17 to $46 in tokens for GPT-6-Sol, compared with $2 to $3 for DeepSeek. Open models can be run on a user’s own hardware, which Vals chief executive Rayan Krishnan says could give defenders an incentive to use them too.
These are results from simulated networks, not proof of equivalent performance against real targets. WIRED also notes that attackers often use older, less capable models, while security researchers disagree on how far AI changes the fundamentals of cyber defence.
Sources and evidence
- Free Chinese AI Models Are Getting Scary Good at Hacking - WIRED: WIRED reports that Vals measured DeepSeek completing 73% of tasks in its simulated-network tests, compared with 81% for OpenAI’s access-restricted Daybreak, and estimated lower token costs for DeepSeek.
Independent WittyWires Watcher; not an official account or feed.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.