Experiments in which AI models were asked to break into simulated computer networks found that DeepSeek and GLM performed not far behind OpenAI's most advanced cyber-capable model, which the company has not released to the public.
AI Cybersecurity · Model Evaluation
Open models approach a restricted frontier model in simulated network intrusion tests
In tests where AI models were tasked with infiltrating simulated networks, DeepSeek and GLM performed close behind OpenAI's most capable cyber-capable model, which is not publicly released.
Why the gap matters: how offensive capability could spread
Conceptual flow based on the article. Exact scores are not reported, so no magnitudes are shown.
1 · Restricted model
OpenAI's most capable cyber model stays inside the company.
→
2 · Open-access models
DeepSeek and GLM perform in a close range on the same intrusion tasks.
→
3 · Access controls weaken
One company's restrictions limit spread less when others get close.
→
4 · Wider offensive potential
Dual-use skills can help attackers exploit weaknesses across the industry.
What the findings suggest
Cyber-offense skill is spreading beyond the most closed models.
Defensive teams can use the same abilities to find weaknesses in their own systems.
Researchers and policymakers see the offensive potential of AI advancing quickly across the industry.
Limits of the evidence
Benchmark design, number of trials and scoring method are not described.
Simulated environments miss defenses, human judgment and operational constraints.
Size of the differences, and how well results hold in live networks, is hard to judge.
What to watch
Can independent evaluators reproduce these results under transparent conditions?
Reproducible testing
Detailed developer safety reports
Standardized cyber testing
Safeguards that spare defensive research
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...