--- id: "2026-09-17-caisi-glm-5-3-cyber-assessment" url: "https://postcutoff.com/e/2026-09-17-caisi-glm-5-3-cyber-assessment/" as_of: "2026-11-22T13:32:00+01:00" date: "2026-09-17" date_precision: day category: policy-safety importance: 4 confidence: high status: [Confirmed] sources: 1 editor: human editor human_review: null version: "2026-10-04" --- As of: 2026-11-22 13:32 CET. Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/ Canonical page: https://postcutoff.com/e/2026-09-17-caisi-glm-5-3-cyber-assessment/ # NIST CAISI: GLM-5.3 is the most cyber-capable open-weight model yet, but trails the US frontier by about four months On Sept 17, 2026 NIST's Center for AI Standards and Innovation (CAISI) published an assessment of Z.ai's GLM-5.3. It called GLM-5.3 "the most cyber-capable open-weight model released to date", ahead of Kimi K3, the previous best Chinese model. It also found the model's cyber capabilities "significantly lower than those of current U.S. frontier models", about four months behind them (e.g. SEC-Bench Pro 40.4% vs 90.2%). ## Key facts - Published by NIST CAISI on 2026-09-17; GLM-5.3 was announced 2026-08-14 and its weights were released about two weeks later - SEC-Bench Pro: GLM-5.3 40.4% vs 90.2% for the best US frontier model - ExploitBench: 61.1% (9.8/16) vs 100.0% (16.0/16) - ExploitGym: 9.4% (47/498) vs 44.4% (223/502) - CAISI OSS-Fuzz: 7.7% (23/297) vs 23.2% (69/297) - Method: models run as agents in a ReAct harness with bash and python; turn limits 200 (SEC-Bench Pro), 300 (ExploitBench), 200 (ExploitGym), 300 (OSS-Fuzz); US models tested with cyber safeguards disabled - CAISI estimate: GLM-5.3 lags the US frontier's cyber capability by about four months; it beats Kimi K3, the previous best PRC model ## What happened The US government's AI evaluation center (CAISI, part of NIST) ran GLM-5.3 and US frontier models through four agentic offensive-security benchmarks: SEC-Bench Pro, ExploitBench, ExploitGym and its own OSS-Fuzz task set. All models ran in the same ReAct agent harness with bash and python tools, and the US models had their cyber safeguards switched off. GLM-5.3 scored higher than any earlier open-weight model, including Moonshot's Kimi K3. On every benchmark it scored well below the best US model, and CAISI put the gap at about four months of progress. ## Why it matters This is the US government's own measurement of the open-weight cyber gap, published twelve days before Anthropic's Frontier Red Team report on the same model. The two reports frame the result differently. CAISI stresses that the US frontier is still clearly ahead. Anthropic stresses that GLM-5.3 nearly matched Claude Mythos Preview on its hardest exploit tasks and that its safeguards can be removed. Read together, they suggest that capabilities the US frontier had a few months earlier are now available to anyone who downloads the weights. ## Your AI and this story - GPT-6.2 Astra (training cutoff May 2026): 109 days after its cutoff - Claude Opus 5.5 (training cutoff June 2026): 79 days after its cutoff - Gemini 4 Hydrogen (training cutoff March 2026): 170 days after its cutoff - Grok 4.7 (training cutoff May 2026): 109 days after its cutoff ## Sources 1. [NIST: CAISI's Assessment of Z.ai's GLM-5.3 Cyber Capabilities](https://www.nist.gov/news-events/news/2026/09/caisis-assessment-zais-glm-53-cyber-capabilities) (nist.gov, official) ## Changes - 2026-10-03 (filed): Created ## Related - 2026-09-29: [Anthropic: open-weights GLM-5.3 nearly matches Mythos Preview at exploit development](https://postcutoff.com/e/2026-09-29-anthropic-glm-5-3-spread-of-cyber-capabilities/index.md) - 2026-09-01: [Hugging Face disables an abliterated GLM-5.3 repo branded "for offensive cyber"](https://postcutoff.com/e/2026-09-01-huggingface-disables-offensive-cyber-glm-5-3/index.md) - 2026-08-14: [Zhipu (Z.ai) releases GLM-5.3, top open-weights coding/agent model](https://postcutoff.com/e/2026-08-14-zhipu-glm-5-3/index.md) - 2026-07-16: [Moonshot AI releases Kimi K3, a 2.8T-parameter open-weights multimodal model](https://postcutoff.com/e/2026-07-16-moonshot-kimi-k3/index.md)