Cantina发布apex-flash-1安全模型,GRPO微调GLM-5.3
Can an Open Model Do Security Research? Cantina’s apex-flash-1 Solves 40 of 60 Held-Out Bug Tasks
开源安全垂直模型的重大进展,GRPO微调路径与极低的推理成本对比极具参考价值,建议关注其工程落地细节。
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.
Cantina Security 与 Yeta Labs 联合发布了 apex-flash-1,这是一个专为漏洞研究训练的开放权重模型。它是基于 Z.ai 的 GLM-5.3-Flash 进行强化学习微调的版本,在 Hugging Face 上以 MIT 许可证发布。
Is it deployable? Yes, the MIT weights serve on vLLM, SGLang or Transformers, but BF16 needs roughly 640 GB of GPU memory.
它可以部署吗?可以,MIT 权重的模型支持 vLLM、SGLang 或 Transformers,但 BF16 精度大约需要 640 GB 的 GPU 显存。
What Cantina Built
Cantina 构建了什么
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.
根据其在 Hugging Face 上的 safetensors 元数据,apex-flash-1 总参数量为 321.3B。GLM-5.3-Flash 基础模型是一个混合专家(MoE)模型,拥有 18B 个激活参数。
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.
Cantina 使用 GRPO 结合 rank-256 LoRA 以及选择性全参数训练对其进行训练。数据涵盖了由 50 个真实漏洞案例构建的 150 个任务。
Each case appears in 3 variants: guided whitebox, focused whitebox and focused blackbox.
每个案例包含 3 种变体:引导式白盒、聚焦式白盒和聚焦式黑盒。
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.
授权、身份和范围缺陷占案例总数的 72%。会计和数值精度错误占 18%。其余部分涵盖时间验证、业务规则和 SSRF(服务器端请求伪造)。
As per the model card on HF, RL rollouts ran inside the Codex agent harness on production-like software and protocol environments.
根据 HF 上的模型卡片说明,RL rollout 是在类似生产环境的软件和协议环境中,通过 Codex agent harness 运行的。
Benchmark Results
基准测试结果
Cantina evaluated 60 tasks from 20 held-out vulnerability cases. Each model ran the set once, with costs estimated from provider pricing.
Cantina 对来自 20 个未参与训练的漏洞案例的 60 个任务进行了评估。每个模型对该任务集运行一次,成本根据供应商定价估算。
- 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
- apex-flash-1:解决 40/60 个任务(pass@1 通过率 66.7%),成本约 $2.38
- GLM-5.3-Flash(基础版):解决 36/60 个任务(通过率 60.0%),成本约 $4.56
- Claude Opus 5 High:解决 43/60 个任务(通过率 71.7%),成本约 $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.
Opus 多解决了 3 个任务,但每次运行的成本高出约 31 倍。这意味着 apex-flash-1 每解决一个任务的成本约为 $0.06,而 Opus 为 $1.74。这些是公司在内部基准测试中报告的数据。
A Worker Model, Not an Orchestrator
作为工作模型,而非编排器
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.
Cantina 将 apex-flash-1 定位为由更大模型编排的工作模型。模型卡片列出的目标技能包括代码阅读、工具使用、漏洞利用开发和验证。
An experimental apex-flash-1-abliterated variant ships with modified refusal behavior. It was not separately evaluated.
一个实验性的 apex-flash-1-abliterated 变体附带了修改后的拒绝行为。该变体未单独进行评估。
Cantina’s rationale is that defenders need capable models they can run and control locally.
Cantina 的理由是,防御者需要能够本地运行和控制的可信模型。
Interactive Explainer
交互式解释器
How It Compares
对比情况
| Feature | apex-flash-1 | Aikido Altar-1 | Cisco Foundation-Sec-8B-Reasoning | GLM-5.3-Flash |
|---|---|---|---|---|
| Developer | Cantina Security + Yeta Labs | Aikido Security | Cisco Foundation AI | Z.ai |
| Base model | GLM-5.3-Flash | GLM-5.3 (pruned) | Llama 3.1 8B | Own pretraining |
| Size | 321.3B total, BF16 | 328 GB, INT4 (W4A16) | 8B | 320B total, 18B active |
| License | MIT | Inherits GLM-5.3 license | Custom (see NOTICE.md) | MIT |
| Security method | GRPO RL on 50 real vulnerability cases | Expert pruning (REAP) + quantization | Instruction tuning + RLHF on security QA | General-purpose base |
| Primary use | Agentic vuln research worker | Air-gapped autonomous pentesting | SOC triage and threat defense | General coding and agents |
| Hardware | Multi-GPU node (~642 GB BF16 weights) | 4x H200 with vLLM | Single GPU | Multi-GPU node |
| Published security result | 66.7% pass@1, 60 tasks | 60.4% recall, 32-CVE internal set | Cisco-reported security benchmarks | 60.0% on Cantina’s set |
| 特性 | apex-flash-1 | Aikido Altar-1 | Cisco Foundation-Sec-8B-Reasoning | GLM-5.3-Flash |
|---|---|---|---|---|
| 开发者 | Cantina Security + Yeta Labs | Aikido Security | Cisco Foundation AI | Z.ai |
| 基础模型 | GLM-5.3-Flash | GLM-5.3 (剪枝) | Llama 3.1 8B | 自有预训练 |
| 规模 | 总计 321.3B,BF16 | 328 GB,INT4 (W4A16) | 8B | 总计 320B,18B 活跃 |
| 许可证 | MIT | 继承 GLM-5.3 许可证 | 自定义(见 NOTICE.md) | MIT |
| 安全方法 | 在 50 个真实漏洞案例上进行 GRPO RL | 专家剪枝 (REAP) + 量化 | 安全 QA 上的指令微调 + RLHF | 通用基础模型 |
| 主要用途 | 代理式漏洞研究工作者 | 气隙环境自主渗透测试 | SOC 分诊与威胁防御 | 通用编码与智能体 |
| 硬件 | 多 GPU 节点(约 642 GB BF16 权重) | 4x H200 配合 vLLM | 单 GPU | 多 GPU 节点 |
| 发布的安全结果 | 66.7% pass@1,60 个任务 | 60.4% 召回率,32-CVE 内部集 | Cisco 报告的安全基准 | Cantina 数据集上为 60.0% |
Sources: Cantina, Aikido, Cisco, Hugging Face model cards. © Marktechpost
来源:Cantina、Aikido、Cisco、Hugging Face 模型卡片。© Marktechpost
Key Takeaways
关键要点
- 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.
- apex-flash-1 是一个基于 MIT 许可证的 321.3B 开源权重安全模型。
- 在 50 个真实漏洞案例上进行 GRPO 训练生成了 150 个任务。
- 其 pass@1 得分为 66.7%,而 Claude Opus 5 High 为 71.7%。
- 其 60 个任务的运行成本约为 2.38 美元,而 Opus 为 74.68 美元。
- BF16 需要多 GPU 节点;社区已有 4-bit 移植版本。
Check out the Technical details. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
查看技术细节。所有功劳归于该项目的研究者。也欢迎我们在 Twitter 上关注我们,别忘了加入我们有 15 万+成员的 ML SubReddit 并订阅我们的新闻通讯。等等!你在 Telegram 上吗?现在你也可以加入我们的 Telegram 群组。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力