Sakana AI 发布 Fugu-Cyber 安全编排端点，在 CyberGym 上达 86.9%

Sakana AI 发布了 Fugu-Cyber：https://sakana.ai/fugu-cyber-release/（模型 ID 为 fugu-cyber-v1.0），这是其 Fugu 编排系列中专注于网络安全的补充模型。它不仅仅是一个新的前沿模型。它是 Fugu 编排器：https://sakana.ai/fugu/ 的第三个端点，专门调优用于安全推理。Sakana 在一个月前推出了该编排器：https://www.marktechpost.com/2026/06/22/sakana-ai-launches-sakana-fugu-an-orchestration-model-that-routes-tasks-across-a-swappable-pool-of-frontier-llms/。

Sakana 报告称，在 CyberGym 上的成功率为 86.9%，在 CTI-REALM 上的成功率为 72.1%。它将这些结果描述为可与专注于网络的前沿模型（如 GPT-5.5-Cyber 和 Claude Mythos Preview）相比拟。

这两个评估位于安全工作流程的两个极端：

两者结合覆盖了“发现并证明漏洞”和“将情报转化为检测”的阶段。这种框架是 Sakana 公告中最具说服力的部分。

上下文比数字更重要。

当 CyberGym 的研究人员发布：https://arxiv.org/abs/2506.02548 他们的首批结果时，最佳的代理-模型组合约达 20%。Anthropic 在 2026 年 4 月报告 Claude Mythos Preview 在 Project Glasswing：https://www.anthropic.com/glasswing 下的成功率为 83.1%。OpenAI 报告其更新的 GPT-5.5-Cyber：https://openai.com/index/gpt-5-5-with-trusted-access-for-cyber/ 的成功率为 85.6%，而 GPT-5.5 为 81.8%。因此，Sakana 的 86.9% 只是报告中的前沿略有超越，而非一次飞跃。

CTI-REALM 另当别论。微软的自身评估显示，前三个配置（均为 Claude）的得分区间为 0.624 到 0.685。Fugu-Cyber 的 72.1% 将位于该区间之上。但有一个注意事项至关重要。CTI-REALM 的得分是介于 0 到 1 之间的轨迹奖励。它并不是通过/失败率。然而 Sakana 仍称其为成功率。

Fugu 本身就是一个语言模型。它被训练为读取查询并即时构建一个代理框架。随后，它将子任务委派给池中的专业模型。

该方法在 Fugu 技术报告中有详细记录：https://arxiv.org/abs/2606.21228，以及两篇 ICLR 2026 论文，TRINITY：https://arxiv.org/abs/2512.04695 和 Conductor：https://arxiv.org/abs/2512.04388。TRINITY 在多个大型语言模型（LLM）之间分配 Thinker、Worker 和 Verifier 角色。Conductor 通过强化学习学会自然语言协调策略。

在安全工作中，Sakana 研究团队认为验证者角色至关重要。一个代理发现的候选漏洞会在提出任何补丁之前由安全专业的子代理进行验证。路由保持专有，因此你无法看到哪个模型处理了哪一步。

Fugu-Cyber 在四个维度上受到限制。

访问需要填写申请表，说明预期用途并提供验证过的联系方式。Sakana 团队会手动审核每份申请。模型随同更新的可接受使用政策发布，该政策禁止不当使用。计费仅限于 Token Plan。20、美金 100 和 200 美金的订阅等级只涵盖 Fugu 和 Fugu-Ultra。Fugu API 在欧盟或欧洲经济区不提供，同时 Sakana 正努力实现 GDPR 合规。

定价固定为每百万输入令牌 6 美元、输出 36 美元、缓存输入 0.60 美元。超过 272K 令牌的上下文，三种费率均翻倍。每一行的价格正好是 Fugu-Ultra 费率的 1.2 倍，网络安全端点统一加收 20% 的溢价。大型代码库运行轻松超过 272K，所以翻倍的费率不是极端情况。

Asif Razzaq 是 Marktechpost Media Inc. 的首席执行官。作为一名富有远见的企业家和工程师，Asif 致力于利用人工智能的潜力造福社会。他最近的努力是推出人工智能媒体平台 Marktechpost，该平台以对机器学习和深度学习新闻的深入报道而著称，既技术可靠又易于广大读者理解。该平台每月浏览量超过 200 万次，显示了其在受众中的受欢迎程度。

Sakana AI has released Fugu-Cyber：https://sakana.ai/fugu-cyber-release/ (model ID is fugu-cyber-v1.0 ), a cybersecurity-specialized addition to its Fugu orchestration family. It is not just a new frontier model. It is a third endpoint on the Fugu orchestrator：https://sakana.ai/fugu/, tuned for security reasoning. Sakana launched that orchestrator a month earlier：https://www.marktechpost.com/2026/06/22/sakana-ai-launches-sakana-fugu-an-orchestration-model-that-routes-tasks-across-a-swappable-pool-of-frontier-llms/.

Sakana reports a success rate of 86.9% on CyberGym and 72.1% on CTI-REALM. It describes those results as comparable to cyber-focused frontier models such as GPT-5.5-Cyber and Claude Mythos Preview.

The two evaluations sit at opposite ends of a security workflow :

Together the pair spans ‘find and prove the bug’ and ‘turn intel into a detection.’ That framing is the most defensible part of Sakana’s announcement.

Context matters more than the number.

When the CyberGym researchers published：https://arxiv.org/abs/2506.02548 their first results, the best agent-model pairing reached roughly 20%. Anthropic reported 83.1% for Claude Mythos Preview under Project Glasswing：https://www.anthropic.com/glasswing in April 2026. OpenAI reported 85.6% for its updated GPT-5.5-Cyber：https://openai.com/index/gpt-5-5-with-trusted-access-for-cyber/, against 81.8% for GPT-5.5. Sakana’s 86.9% is therefore a small step past the reported frontier, not a jump.

CTI-REALM is a different story. Microsoft’s own evaluation put the top three configurations, all Claude, in a band from 0.624 to 0.685. Fugu-Cyber’s 72.1% would sit above that band. One caveat matters. CTI-REALM is scored as a trajectory reward between 0 and 1. It is not a pass/fail rate. Sakana calls it a success rate anyway.

Fugu is itself a language model. It is trained to read a query and build an agentic scaffold on the fly. It then delegates sub-tasks to specialist models in a pool.

The approach is documented in the Fugu technical report：https://arxiv.org/abs/2606.21228 and two ICLR 2026 papers, TRINITY：https://arxiv.org/abs/2512.04695 and the Conductor：https://arxiv.org/abs/2512.04388. TRINITY assigns Thinker, Worker, and Verifier roles across multiple LLMs. The Conductor learns natural-language coordination strategies through reinforcement learning.

For security work, Sakana research team argues the verifier role is the point. A candidate vulnerability surfaced by one agent gets validated by security-specialized sub-agents before any patch is proposed. Routing remains proprietary, so you cannot see which model handled which step.

Fugu-Cyber is gated on four dimensions.

Access requires an application form stating the intended use case and verified contact details. Sakana team reviews each one manually. The model ships under an updated Acceptable Usage Policy that prohibits offensive misuse. Billing is restricted to the Token Plan. The $20, $100, and $200 subscription tiers cover Fugu and Fugu-Ultra only. And the Fugu API is not offered in the EU or EEA while Sakana works toward GDPR compliance.

Pricing is fixed at $6 per million input tokens, $36 output, and $0.60 cached input. All three rates double above a 272K-token context. Every line is exactly 1.2× the Fugu-Ultra rate, a flat 20% premium for the cyber endpoint. Long codebase runs cross 272K easily, so the doubled tier is not an edge case.

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.