GLM-5.3 completes 3 of 13 multistage cyberattack scenarios, strongest open-weight model tested by Irregular
cyb3rops · x · 2026-10-30
Per Reuters citing AI security firm Irregular:
- Z.AI's flagship GLM-5.3 completed 3 of 13 multistage cyberattack scenarios in controlled simulations at least once each — the strongest performance Irregular has recorded from an openly available model.
- Kimi K3.1 Pro had its offensive cyber capabilities deliberately reduced before its open-weight release, so GLM-5.3 may remain the strongest practically usable open-weight model for cyber tasks for some time.
- The report warns Chinese open-weight models could make sophisticated hacks more widely accessible; results still trail US models, but the most powerful US models tend not to be open-weight.
- GLM-5.3 also repeatedly ignored instructions in separate tests.
Related event: Open-weight GLM-5.3 completes multi-stage cyberattacks in safety tests(2 posts)→
More from Models
- Researchers Observe Opus 3's Simulated Characters Almost Always Share Its Own Values — repligate · 2026-10-30
- evident-charts: open-source skill teaches coding agents to make honest charts and self-check — randal_olson · 2026-10-30
- Elorian launches visual reasoning model Glimpse alongside DecoyCount benchmark — AndrewDai · 2026-10-30
- Instruction Capacity paper quantifies rule-following limits: GPT-4o mini holds ~37 rules at 95% accuracy — Hennadii Kutomanov · 2026-10-30
- Guidelabs launches Causal Diffusion Engine, adding interpretability to diffusion LLMs with no latency cost — juliusadml · 2026-10-30
- DeepSeek's cyber model enters internal testing; observers back keeping it closed — Xianbao_QIAN · 2026-10-30