By Roman Bodnarchuk
10XAI.news | Special Series 5 of 14 | Research checked October 18, 2026
If your security plan assumes capable AI is available only through a few tightly controlled services, it is time to revisit the assumption.
What the research says
Epoch AI’s October 16 research listing reports that open-weight models trail the closed frontier by approximately six to seven months on its Cyber ECI. It identifies Z.ai’s GLM-5.3 as the leading open-weight model on that index. Epoch’s methodology page notes that the Cyber ECI includes cybersecurity benchmarks outside its general index and is updated less frequently. [1][2]
That is a comparison on a defined set of evaluations. It is not a statement that every open model can compromise any organization, or that incidents have increased by a corresponding amount.
My judgment
Use the finding to update your assumptions and test your controls.
For an enterprise leader, the immediate work is specific. Know which systems are exposed, who can grant an agent access, and whether an unusual action produces an observable record.
There is a useful distinction between model capability and operational access. A capable model with no credentials faces a different situation from an ordinary automation tool holding excessive permissions. Your own deployment decisions can change that situation.
Ask your security lead for a bounded review of one existing AI workflow:
- What information can the assistant read?
- Which systems can it change?
- Which credentials or sessions make those actions possible?
- Can a document or web page influence the assistant’s next action?
- What would alert an operator to an unexpected action?
- How quickly could the organization revoke access and recover?
The answers should describe the actual configuration. A vendor’s general security page cannot tell you whether your team connected an overly broad account last Tuesday.
This is work for authorized defensive testing. Keep evaluation inside systems and data you are permitted to assess.
Your 10X action
Pick one assistant that already has tool access. Have its owner produce a one-page access map and demonstrate how that access can be revoked.
Copy this prompt
Help me document this authorized AI workflow: [description]. List data sources, tool permissions, credential owners, approval steps, logs, and recovery options. Mark every permission we have not verified. Propose a defensive tabletop exercise with synthetic data and no external targets. Explain which benchmark claims are relevant and which do not establish anything about our deployment.
Reply with the permission you would check first in your own organization.
Roman Bodnarchuk
10XAI.news