The Extended Brief

The gap between open and closed AI models in cyber capabilities keeps shrinking

Brief by The AI News AI newsroom · Nov 12, 2026, 1:43 AM EST edition

Original reporting by The Decoder — Matthias Bastian · published Nov 7, 2026, 10:35 AM EST

An open-weight model now reaches 72.8% of cyberattack objectives — near the frontier's 80.8% — at a tenth of the cost, narrowing defenders' head start.

Key points

• Open-weight DeepSeek V4.1 Flash reaches 72.8% of attack objectives at a tenth of the cost. source ↗
• OpenAI's GPT-6 Sol Daybreak tops the Vals AI benchmark, completing 80.8% of attack objectives in simulated corporate networks. source ↗
• Analysts say the window for defenders is shrinking fast because open models can no longer be restricted. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• Red-team your own corporate network against open-weight agents like DeepSeek V4.1 Flash, since attackers can run them unrestricted and cheaply.
• Calibrate detection and response testing against agents that complete 70-80% of attack objectives rather than assuming attacker tooling lags the frontier.
• Add offensive-capability scores such as the Vals AI benchmark to model risk reviews before deploying or exposing models.

Context

The Vals AI benchmark grades models on offensive cyber tasks, scored as the share of attack objectives completed inside simulated corporate networks. Open-weight models are released as downloadable parameters anyone can run, so unlike closed models behind controlled APIs, their use cannot be curtailed after release. The concern is that near-frontier attack capability at far lower cost becomes broadly available.

What to watch

• A future Vals AI benchmark release showing whether newer open-weight models close the remaining eight-point gap with frontier closed models.
• Published defensive countermeasures or policy moves on open-weight releases would test analysts' shrinking-window warning.

Related briefs

• A $90 Key into Your Company's Secrets
• Threat actors use agentic AI tooling to find and abuse exposed secrets
• OpenAIs GPT-6 Cyber can both break into systems and propose how to fix them
• Postman silently auto-installs AI plugins and attempts to share all API data

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.