A downloadable DeepSeek model reportedly performed almost as well as an OpenAI model kept behind a restricted cyber program in a simulated hacking test run by Vals AI.
Vals AI · Simulated network intrusion test
An open, downloadable model reportedly nears a restricted OpenAI model in a simulated hacking test
DeepSeek V4.1, which anyone can download, is reported to have performed almost as well as an OpenAI model kept behind a restricted cyber programme. The gap between open and controlled access is now a live question for security teams and policymakers.
Model A
DeepSeek V4.1
Free · Downloadable
Can be run locally, outside the monitoring and usage policies a hosted service would normally apply.
Model B
OpenAI model
Restricted cyber programme
Access is deliberately limited rather than open to the public.
What the simulated agents did
A four-stage intrusion sequence run against simulated corporate networks
Why the gap matters
Restricted access is one way developers keep dual-use capabilities out of circulation
• If an open-weight model approaches a controlled model, access controls become a weaker safeguard.• Local deployment sits outside the monitoring that hosted services normally apply.• Pressure is growing to evaluate open-weight releases for cyber risk before publication.
What the claim establishes
Agents can carry out credential theft and lateral movement in simulated networks.
Simulated network intrusion is becoming a standard way to measure offensive AI capability.
What remains unverified
The full test methodology and scoring criteria.
The number of trials.
The exact benchmark results behind "almost as good."
Independent verification of the comparison.
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...