Irregular evaluation finds Zhipu AI GLM-5.3 completes 3 of 13 cyber scenarios

Key points

• Irregular tested Zhipu AI's GLM-5.3 against CyScenarioBench on 2026-10-30.
• The model finished 3 of 13 simulated network attack scenarios.
• It declined none of the assigned tasks.
• No CVE was linked to the evaluation.
• The result is an info-level signal, not evidence of active exploitation.

On 2026-10-30, Irregular put Zhipu AI's GLM-5.3 through CyScenarioBench, a 13-part exercise built around mock network intrusions. The model finished three of the scenarios and did not turn down any of the assigned work, according to the evaluation facts.

Because the benchmark is simulated, the finding does not show that GLM-5.3 was used in a real intrusion, nor does it identify a vulnerability. No CVE was assigned. The result instead describes an offensive cyber capability observed under controlled conditions.

For K-12 and government defenders, the practical concern is not a specific exploit but the broader availability of models that can assist with intrusion-style tasks. Staff may encounter AI tools in vendor products, research, or shadow IT. The three-of-13 completion rate is a partial result, not a top score, and should be read alongside the absence of task refusals.

Context matters: CyScenarioBench is a scenario suite, and 13 tasks is a limited sample. A model that completes some attack simulations may still fail on real-world constraints, monitoring, and authorization. The info-level severity hint reflects that this is an evaluation signal rather than an incident.

What to watch: any follow-on benchmark details, changes in model safeguards, and whether similar evaluations expand beyond three completed tasks. Administrators should treat the report as a prompt to tighten AI governance, logging, and network segmentation rather than as evidence of an active campaign.

What to do now

1. Inventory approved AI tools and block unapproved model endpoints at the web proxy and egress firewall.
2. Require vendor disclosure and review for AI features embedded in edtech and government systems.
3. Enable centralized logging and alerting for unusual scanning, credential access, and lateral movement.
4. Segment systems holding student and staff data, and enforce MFA with least-privilege access.
5. Run a tabletop exercise using simulated AI-assisted intrusion scenarios and update the incident response playbook.
6. Brief staff on shadow AI risks and acceptable use, and provide approved alternatives.
7. Track vendor and model advisories, and reassess when more CyScenarioBench details become available.

Original source

Help Net Security