GLM-5.3 nearly matches Claude at exploiting bugs; $20 in tokens found a Chrome flaw
DeepLearningAI · x · 2026-10-04
In this week's letter, Andrew Ng analyzes Anthropic's evaluation of open-weight GLM-5.3's cyber capabilities: on a subset of ExploitBench exploit tasks, GLM-5.3 solved 12% versus closed-weight Claude Mythos's 14% at comparable token cost (GLM's own full-benchmark numbers show a wider gap: 54.4% vs 78.0%).
- Falling attack costs: Anthropic found $20.40 of GLM-5.3-Flash tokens sufficed to find a recently disclosed Google Chrome vulnerability, and guardrail-weakened open models are easy to obtain.
- Defenders benefit: teams without access to Mythos can still use highly capable open models to harden defenses.
- Ng's take: AI cyber risk is an engineering problem, not a reason to panic — defenders hold the long-term edge and should ramp up defenses now.
Related event: Open-source GLM-5.3 nearly matches Claude in exploit capability(2 posts)→
More from Models
- Mystery Model 'Gestalt' Tops New Frontier Design Benchmark by 214 Points — soleio · 2026-11-19
- Gemini 4 Neon arrives days after Hydrogen: fast reasoning at $3/$15 per 1M tokens — TorturedPoet30 · 2026-11-19
- Screenshot confirms Gemini 4 Neon launch at $3/$15 per million tokens — Recoil42 · 2026-11-19
- mrfakename releases abliterated GLM-5.3-Flash and Qwen3.8-Flash-Next — mrfakename0 · 2026-11-19
- Physician: AI competition now measured in weeks, GPT-6 versions 7 days apart — ramos_casals · 2026-11-19
- micro1's flow qc model screens training data for benchmark leaks before release — Exp_Mark · 2026-11-19