Can an Open Model Do Security Research? Cantina’s apex-flash-1 Solves 40 of 60 Held-Out Bug Tasks
https://www.marktechpost.com/2026/10/04/can-an-open-model-do-security-research-cantinas-apex-flash-1-solves-40-of-60-held-out-bug-tasks/📌 開源安全模型對決 Opus：成本僅 1/31,漏洞任務解 40/60
TL;DR：Cantina 開源 apex-flash-1,專攻漏洞研究,成本效益數據直接攤在你面前。
當大家還在爭論「要不要讓 LLM 碰生產環境的程式碼審查」時,Cantina Security 已經直接把一個專門找漏洞的模型放上 Hugging Face,還附上跟 Claude Opus 的正面對決成績單。
🤔 防禦方也需要能本機掌控的模型
Cantina Security 與 Yeta Labs 合作發布 apex-flash-1,一個專為漏洞研究訓練的開放權重模型,以 MIT 授權釋出。Cantina 的立場很直接:防禦方需要能在自己環境中執行、控管的強力模型,而不是只能透過 API 呼叫外部黑盒。
🧩 GLM-5.3-Flash 底座,GRPO 強化學習微調
apex-flash-1 是以 Z.ai 的 GLM-5.3-Flash 為底座,透過強化學習微調而成。根據 Hugging Face 的 safetensors metadata,模型總參數量為 321.3B,底座 GLM-5.3-Flash 本身是 Mixture-of-Experts 架構,啟用參數為 18B。Cantina 使用 GRPO 搭配 rank-256 的 LoRA,再加上選擇性的全參數訓練。
訓練資料來自 50 個真實漏洞案例,衍生出 150 個任務,每個案例各有三種變體:guided whitebox、focused whitebox、focused blackbox。案例類型分布上,授權、身分驗證與範圍（scope）相關的漏洞佔 72%,帳務與數值精度類漏洞佔 18%,其餘則是時間驗證、業務邏輯與 SSRF。依模型卡說明,RL rollout 是在 Codex agent harness 中、對接近生產環境的軟體與協定環境下進行的。
📊 40/60 任務,成本只要 Opus 的 1/31
Cantina 用 20 個held-out漏洞案例、共 60 個任務進行評測,每個模型各跑一次,成本依各家供應商定價估算。結果 apex-flash-1 解出 40 個任務,Opus 多解出 3 個,但每次執行成本高出約 31 倍——換算下來,apex-flash-1 每解決一個任務約 0.06 美元,Opus 則約 1.74 美元。需要強調的是,這些是 Cantina 自行在內部 benchmark 上公布的數字。
💡 模型的定位:被更大模型指揮的「工人」
Cantina 把 apex-flash-1 定位為由更大模型orchestrate的worker,模型卡列出的目標技能包括程式碼閱讀、工具使用、漏洞利用開發與驗證。另外還有一個實驗性的 apex-flash-1-abliterated 變體,修改了拒絕回應的行為,但這個版本並未單獨評測。
⚠️ 權重好拿,但硬體門檻不低
MIT 授權的權重可在 vLLM、SGLang 或 Transformers 上部署,但 BF16 精度大約需要 640GB GPU 記憶體,這對多數團隊而言仍是不小的基礎設施投入。此外,所有效能與成本比較都是 Cantina 單方面在內部評測集上得出,尚未見到第三方獨立複現。
🎯 實務啟示
如果你的團隊在做內部紅隊或漏洞研究自動化,apex-flash-1 的成本結構值得關注:用小模型搭配更大模型做 orchestration,可能比全程呼叫頂級閉源模型更划算。但部署前務必評估 640GB 記憶體需求是否合乎團隊現有硬體,以及這份 40/60 的成績單是否能在你自己的漏洞類型上重現。
🔗 來源
- 標題：Can an Open Model Do Security Research? Cantina’s apex-flash-1 Solves 40 of 60 Held-Out Bug Tasks
- 作者／機構：Michal Sutter, MarkTechPost
- 連結：https://www.marktechpost.com/2026/10/04/can-an-open-model-do-security-research-cantinas-apex-flash-1-solves-40-of-60-held-out-bug-tasks/
#SecurityResearch #OpenWeights #LLM #VulnerabilityResearch #ReinforcementLearning #GRPO #MixtureOfExperts #AIforSecurity #HuggingFace #CyberSecurity
原始資料 MarkTechPost · 收集於 2026-10-05
摘要原文
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. 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. Cantina evaluated 60 tasks from 20 held-out vulnerability cases. Each model ran the set once, with costs estimated from provider pricing. 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. Sources: Cantina , Aikido , Cisco , Hugging Face model cards. © Marktechpost 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. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Can an Open Model Do Security Research? Cantina’s apex-flash-1 Solves 40 of 60 Held-Out Bug Tasks appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成