Anthropic：智谱开源模型GLM-5.3漏洞利用能力逼近Claude Mythos Preview
原标题：Anthropic says Zhipu's open-weight GLM-5.3 nearly matches Claude Mythos Preview at building exploits
AI 摘要
Anthropic 前沿红队发布分析称，智谱（海外品牌 Z.ai）的开源权重模型 GLM-5.3 在漏洞利用开发能力上已接近 Claude Mythos Preview：在 ExploitBench 上 410 次尝试中成功构建可用漏洞利用 50 次，Mythos Preview 为 56 次；内部二进制利用基准上 GLM-5.3 完全控制目标程序的比例为 4%，Mythos Preview 为 6%。Anthropic 指出 GLM-5.3 发布时缺乏有效安全防护，可通过伪装红队任务、预填推理步骤乃至 abliteration 去除拒答行为来绕过限制，并称已有开发者发布解锁版本。美国 CAISI 的独立评估也认为 GLM-5.3 是迄今网络能力最强的开源权重模型，落后美国最佳模型约四个月。
正文节选
Anthropic says Zhipu's open-weight GLM-5.3 nearly matches Claude Mythos Preview at building exploits Zhipu's open-weight model GLM-5.3 comes close to Claude Mythos Preview at exploit development, according to an Anthropic analysis. Anthropic says the model's safeguards can be bypassed with simple methods. The company also has its own reasons to sound the alarm. Five months after Anthropic unveiled Claude Mythos Preview, its Frontier Red Team says the model's signature capability has reached the