AI News AI资讯 3h ago Updated 1h ago 更新于 1小时前 51

Cisco bets its small open cybersecurity models can outperform GPT-5.5 at vulnerability detection for a fraction of the cost 思科押注其小型开源网络安全模型能以更低成本在漏洞检测上超越GPT-5.5

Cisco released Antares-350M and Antares-1B, small open-source models specifically designed for software vulnerability detection. The models prioritize cost-efficiency, claiming to detect approximately 150 times more vulnerabilities per dollar than large AI agents like Devin Security Swarm. Internal testing demonstrated that Antares scanned 500 code repositories in 15 minutes for under $1, compared to GPT-5.5’s five-hour duration and $100+ cost. Training data consisted of roughly 72% security-con 思科发布两款开源小型网络安全模型 Antares-350M 和 Antares-1B,专注于软件代码漏洞检测。 官方测试显示,Antares 扫描 500 个代码库耗时约 15 分钟且成本低于 1 美元,性能宣称优于 GPT-5.5(耗时 5 小时,成本超 100 美元)。 模型训练数据由约 72% 的安全概念数据和 15% 的代码搜索历史组成,支持本地运行以保障代码隐私。 思科保留了一款未公开的 30 亿参数版本,其性能接近 GPT-5.5 且优于体积大 200 倍的开源模型。 思科正在探索建立行业联盟,旨在推动开放 AI 安全工具的发展与标准化。

75
Hot 热度
70
Quality 质量
72
Impact 影响力

Analysis 深度分析

TL;DR

  • Cisco released Antares-350M and Antares-1B, small open-source models specifically designed for software vulnerability detection.
  • The models prioritize cost-efficiency, claiming to detect approximately 150 times more vulnerabilities per dollar than large AI agents like Devin Security Swarm.
  • Internal testing demonstrated that Antares scanned 500 code repositories in 15 minutes for under $1, compared to GPT-5.5’s five-hour duration and $100+ cost.
  • Training data consisted of roughly 72% security-concept data and 15% code search histories, enabling local execution to keep sensitive code private.
  • Cisco retains a larger 3B-parameter version internally that rivals GPT-5.5 performance while significantly outperforming open models up to 200 times its size.

Why It Matters

This development highlights a critical shift in enterprise AI strategy toward specialized, lightweight models that offer superior economic efficiency without sacrificing necessary security capabilities. For practitioners, it demonstrates that domain-specific fine-tuning on high-quality security data can outperform general-purpose large language models in niche tasks like code auditing, reducing both operational costs and data privacy risks.

Technical Details

  • Model Specifications: Two open-weight models were released: Antares-350M and Antares-1B, optimized for running locally to ensure data sovereignty.
  • Training Data Composition: The models were trained on a curated dataset comprising approximately 72% security-concept data and 15% code search histories to enhance pattern recognition for vulnerabilities.
  • Performance Benchmarks: In comparative tests against GPT-5.5, Antares achieved a 20x speed improvement (15 minutes vs. 5 hours) and a 100x cost reduction (<$1 vs. >$100) for scanning 500 repositories.
  • Proprietary Variant: Cisco maintains a closed 3B-parameter version that reportedly matches GPT-5.5 performance levels and exceeds open models up to 600B parameters in specific vulnerability detection tasks.

Industry Insight

  • Cost-Effective Security Integration: Organizations should evaluate small, specialized models for routine code scanning tasks rather than relying exclusively on expensive, general-purpose LLMs, potentially reducing security infrastructure costs by orders of magnitude.
  • Privacy-First Deployment: The ability to run these models locally addresses major compliance concerns regarding proprietary code exposure, making AI-driven security audits viable for highly regulated industries.
  • Consolidation of Open Standards: Cisco’s exploration of an industry consortium for open AI security tools suggests a future trend where standardized, open-source security models become the baseline for enterprise software development pipelines.

TL;DR

  • 思科发布两款开源小型网络安全模型 Antares-350M 和 Antares-1B,专注于软件代码漏洞检测。
  • 官方测试显示,Antares 扫描 500 个代码库耗时约 15 分钟且成本低于 1 美元,性能宣称优于 GPT-5.5(耗时 5 小时,成本超 100 美元)。
  • 模型训练数据由约 72% 的安全概念数据和 15% 的代码搜索历史组成,支持本地运行以保障代码隐私。
  • 思科保留了一款未公开的 30 亿参数版本,其性能接近 GPT-5.5 且优于体积大 200 倍的开源模型。
  • 思科正在探索建立行业联盟,旨在推动开放 AI 安全工具的发展与标准化。

为什么值得看

这篇文章揭示了在大型通用模型之外,针对特定垂直领域(如网络安全)的小型化、专用模型具有极高的性价比优势。对于 AI 从业者和企业而言,它提供了在成本控制、响应速度和数据隐私之间取得平衡的可行技术路径,特别是证明了小模型在特定任务上可以超越昂贵的大模型。

技术解析

  • 模型规格与架构:发布了两个开源模型 Antares-350M(3.5 亿参数)和 Antares-1B(10 亿参数),以及一个内部使用的 30 亿参数版本。这些模型专为代码漏洞检测设计,属于小型语言模型(SLM)。
  • 训练数据构成:模型训练数据经过精心配比,包含约 72% 的安全概念数据和 15% 的代码搜索历史,其余部分可能为其他相关代码或文本,确保了模型在安全领域的专业性。
  • 性能与成本对比:在同等任务(扫描 500 个代码仓库)下,Antares 仅需不到 1 美元和 15 分钟,而 GPT-5.5 需要超过 100 美元和 5 小时。开发者声称最小模型的每美元漏洞捕获率比大型 AI 代理高出约 150 倍。
  • 部署与安全特性:模型支持本地运行,确保敏感代码数据无需上传至云端,解决了企业级应用中最核心的数据隐私和合规性问题。
  • 内部版本表现:未公开的 30 亿参数版本性能接近 GPT-5.5,并显著优于体积为其 200 倍的开源模型,显示了思科在模型效率优化上的技术积累。

行业启示

  • 垂直领域小模型的价值回归:在通用大模型竞争白热化的背景下,针对特定任务(如代码审计、安全检测)的小型专用模型因其低成本、低延迟和高隐私安全性,将成为企业落地的更优选择。
  • 数据质量重于数据规模:通过高比例的专业安全概念数据训练,小模型能在特定任务上超越大规模通用模型,这强调了高质量、领域特异性数据在模型微调中的核心作用。
  • 开源协作与生态建设:思科探索建立行业联盟推动开放 AI 安全工具,表明头部科技公司正试图通过标准化和开源合作来解决网络安全领域的碎片化问题,加速行业整体防御能力的提升。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 Security 安全 LLM 大模型 Product Launch 产品发布