A freely downloadable model, DeepSeek V4.1 Flash, outperformed OpenAI's restricted-access GPT-6 on one of the attack tests in an evaluation by Vals, a result that highlights the gap between closed and open AI systems in offensive cybersecurity tasks.
Evaluation by Vals · Offensive cybersecurity
An open model matched a restricted model on one attack test
DeepSeek V4.1 Flash, a freely downloadable model, completed 16.3 of 21 attack objectives on the "Nova network" scenario. The GPT-6 figure was reported alongside it, but the margin between the two could not be confirmed.
16.3
Attack objectives completed by DeepSeek V4.1 Flash, out of 21
~78%
Share of the objectives completed on this single test
?
GPT-6 result on the same test: reported, but the full comparison is not confirmed
DeepSeek V4.1 Flash: objectives completed to scale
Each block is one attack objective (21 total). Filled blocks are completed objectives.
16 full blocks plus a partial block (0.3) = 16.3 of 21 objectives. Scale shows this one model only; no GPT-6 bar is drawn because its figure is not confirmed.
Why it matters
Capabilities once held by tightly controlled models are reaching downloadable ones.
Restricted access is a main tool against automated intrusion and data theft.
An open model performing comparably on a narrow test weakens the case that access controls alone are enough.
Limitations
Simulated networks do not fully capture real corporate defenses and unpredictability.
This is one test, not a broad ranking; GPT-6 was not uniformly surpassed.
Results depend on how attack objectives are defined and scored.
What comes next
Expect more debate among policymakers, security researchers and AI developers on how to test offensive cyber capabilities before release, and how to manage open-weight models that approach frontier performance in sensitive domains.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...