DeepSeek and GLM come close to OpenAI's unreleased cyber model in break-in test

single source· 1 articles · confidence: low · first seen 2026-11-06 18:43 UTC

What this means for you

Nothing to act on yet: no scores, no test date and no method have been published, and the OpenAI model used as the reference point is not available to compare against. For anyone running models with outbound network access, how far this carries over to a live environment is not answered here.

An experiment that asked AI models to break into simulated networks found DeepSeek and GLM close behind OpenAI's most advanced cyber model — one built for security work — according to Wired. That OpenAI model is not public, so there is no openly checkable baseline. The report, published 6 November 2026, gives no scores, no evaluation date, no model versions and no description of the harness (the scaffolding that lets a model take repeated actions). Whether the result holds under other test conditions is unknown.

Key facts

  • ·Wired reported on 6 November 2026 that an experiment asked AI models to break into simulated networks. source
  • ·The models named in the report are DeepSeek and GLM, tested against OpenAI's most advanced cyber model. source
  • ·That OpenAI model is not publicly available, according to the report. source
  • ·The report gives no scores, no evaluation date and no description of the harness used. source

What the sources say

  • Wired AI — Single short report on a break-in experiment; names the models compared but gives no scores or method.

Sources

The original reporting. Follow these — they did the work.

← the wire