--- format: "aidr-story-markdown/v1" id: "f99c64224da7d0b9d6440fd3d1a2556884f4a5908b66f6aae917246a26be8232" canonical_url: "https://aidr.today/f99c6422?lang=en" title: "Abliterated GLM 5.3 Model Debuts for Offensive Cyber, Ranking No. 3 on Terminal Bench 4.0" lang: "en" requested_lang: "en" available_langs: ["en","vi"] translation_fallback: null fallback_fields: [] published_at: "2026-09-01T06:14:53.000Z" category: "Models" topics: ["glm","safety","agent","coding","benchmark"] source_urls: ["https://huggingnews.com/cybersecurity/abliterated-glm-53-model-debuts-for-offensive-cyber-ranking-no-3-on-term-49a58a8d","https://x.com/WesRoth/status/2094610221339738228","https://x.com/emollick/status/2094617053730685054","https://x.com/max_paperclips/status/2094621917441597472"] summary: "A new AI tool called abliterated-model-large-v2 provides offensive cyber, red teaming, and agent testing capabilities that standard models typically refuse to perform. Based on GLM-5.3, the US-hosted model utilizes FP8 precision and features a 1 million context window with zero input or output prompt retention. The model provides 2x the cyber exploitation capability of GLM-5.2 and removes standard safety guardrails to facilitate the simulation of cyber attacks. It is specifically built to perform the offensive security work that commercial models are programmed to block. GLM-5.3 ranks third on the Terminal-Bench 4.0 benchmark, trailing only Opus 5 and Fable. The release is now live, prompting security researchers to advise immediate red teaming of existing systems to identify and patch vulnerabilities before they can be exploited." --- # Abliterated GLM 5\.3 Model Debuts for Offensive Cyber, Ranking No\. 3 on Terminal Bench 4\.0 > [Open the canonical story]() **Published:** 2026-09-01T06:14:53.000Z **Category:** Models **Topics:** glm, safety, agent, coding, benchmark ## Summary A new AI tool called abliterated\-model\-large\-v2 provides offensive cyber, red teaming, and agent testing capabilities that standard models typically refuse to perform\. Based on GLM\-5\.3, the US\-hosted model utilizes FP8 precision and features a 1 million context window with zero input or output prompt retention\. The model provides 2x the cyber exploitation capability of GLM\-5\.2 and removes standard safety guardrails to facilitate the simulation of cyber attacks\. It is specifically built to perform the offensive security work that commercial models are programmed to block\. GLM\-5\.3 ranks third on the Terminal\-Bench 4\.0 benchmark, trailing only Opus 5 and Fable\. The release is now live, prompting security researchers to advise immediate red teaming of existing systems to identify and patch vulnerabilities before they can be exploited\. ## Sources - [Story source]() - [Story source]() - [Supporting source]() - [Supporting source]()