Cantina 发布开源安全模型 apex-flash-1，60 项漏洞任务解出 40 项
Can an Open Model Do Security Research? Cantina's apex-flash-1 Solves 40 of 60 Held-Out Bug Tasks - MarkTechPost
Cantina Security 与 Yeta Labs 发布面向漏洞研究的开源权重模型 apex-flash-1，基于智谱 GLM-5.3-Flash 做强化学习微调，以 MIT 许可在 Hugging Face 发布。该模型总参数 321.3B，用 GRPO 加 rank-256 LoRA 训练，数据来自 50 个真实漏洞案例构建的 150 个任务。在 20 个留出漏洞案例的 60 项任务上解出 40 项（66.7% pass@1），单次运行约 2.38 美元，Claude Opus 5 High 解出 43 项（71.7%）但成本约 74.68 美元；BF16 部署约需 640 GB GPU 显存。
推荐理由
AI 评分 58
开源安全模型在内部基准上以约三十分之一的成本接近闭源前沿模型，读者可据此判断本地化漏洞研究的可行性。
主题
正文
当前显示原文，中文全文译文尚未就绪。
Cantina Security, with Yeta Labs, has released apex-flash-1, an open-weights model trained specifically for vulnerability research. It is a reinforcement learning fine-tune of Z.ai’s GLM-5.3-Flash, released on Hugging Face under the MIT license.
Is it deployable? Yes, the MIT weights serve on vLLM, SGLang or Transformers, but BF16 needs roughly 640 GB of GPU memory.
What Cantina Built
apex-flash-1 has 321.3B total parameters, per its Hugging Face safetensors metadata. The GLM-5.3-Flash base is a Mixture-of-Experts model with 18B active parameters.
Cantina trained it with GRPO using a rank-256 LoRA plus selective full-parameter training. The data covers 150 tasks built from 50 real vulnerability cases.
Each case appears in 3 variants: guided whitebox, focused whitebox and focused blackbox.
Authorization, identity and scope flaws make up 72% of cases. Accounting and numerical precision bugs add 18%. Time validation, business rules and SSRF cover the rest.
As per the model card on HF, RL rollouts ran inside the Codex agent harness on production-like software and protocol environments.
Benchmark Results
Cantina evaluated 60 tasks from 20 held-out vulnerability cases. Each model ran the set once, with costs estimated from provider pricing.
apex-flash-1: 40/60 solved (66.7% pass@1), about $2.38; GLM-5.3-Flash (base): 36/60 solved (60.0%), about $4.56; Claude Opus 5 High: 43/60 solved (71.7%), about $74.68
Opus solved 3 more tasks but cost about 31x more per run. That is roughly $0.06 per solved task for apex-flash-1 versus $1.74 for Opus. These are company-reported numbers on an internal benchmark.
Cantina positions apex-flash-1 as a worker orchestrated by a larger model. The card lists code reading, tool use, exploit development and verification as target skills.
An experimental apex-flash-1-abliterated variant ships with modified refusal behavior. It was not separately evaluated.
Cantina’s rationale is that defenders need capable models they can run and control locally.
apex-flash-1 is a 321.3B open-weights security model under MIT; GRPO training on 50 real vulnerability cases produced 150 tasks; It scored 66.7% pass@1 versus 71.7% for Claude Opus 5 High; Its 60-task run cost about $2.38 versus $74.68 for Opus; BF16 needs a multi-GPU node; community 4-bit ports exist
Michal Sutter
Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.
Michal SutterGoogle Research Moves Federated Learning Into TEEs: Gboard Now Trains With Externally Verifiable Differential Privacy; Michal SutterDeepSeek Harness v0.2 Brings Official Desktop Apps to Its Open-Source Agent Harness; Michal SutterPrime Intellect Launches Prime Inference: Serverless and Reserved Serving for Frontier Open Models; Michal SutterMicrosoft AI Releases MAI-Transcribe-2-Streaming: #1 Real-Time Speech-to-Text Model on Artificial Analysis