Anthropic 已开始扩大其用于查找漏洞的 Mythos 模型的访问权限，允许 Cisco 和 IBM 将其纳入各自的安全产品中。
Anthropic has started extending access to its bug-finding Mythos model by allowing Cisco and IBM to make it part of their security products.
Cisco 今日在澳大利亚墨尔本举行的一次活动上宣布了这一集成，透露其 AI Defense 套件中的红队工具将能够调用 Mythos 来测试客户的部署。AI Defense 套件用于查找 AI 模型、智能体及其驱动的应用程序中的弱点和不当活动。“从宏观层面来看，AI Defense 有能力审视模型并攻破它们，”Cisco 产品组合战略与营销高级副总裁 Jeff Schultz 表示。
Cisco announced its integration today at an event in Melbourne, Australia, where it revealed that the red teaming tools in its AI Defense suite, which looks for weaknesses and inappropriate activity in AI models, agents, and applications they power, will be able to call Mythos to test customers’ setups. “At a very high level, AI Defense has the ability to look at models and break them,” said Jeff Schultz, Cisco’s senior vice president for portfolio strategy and marketing.
他称该工具现在“由 Mythos 强力驱动”，将识别出更多危险，随后由 AI Defense 进行处理。AI Defense 已经使用基于 AI 的红队测试来发现缺陷。Mythos 将成为另一个选项。Cisco 将在用户现有许可证下提供该功能，无需额外费用。
He said the tool is now “super-charged by Mythos” and will identify more danger that AI Defense will then address. AI Defense already uses AI-powered red teaming to find flaws. Mythos will become another option. Cisco will make it available under users’ existing licenses, without additional cost.
Anthropic 亚太区应用 AI 负责人 Natalie Mead 表示，公司有一长串想要使用 Mythos 的买家，与 Cisco 合作将使满足这一需求变得更容易。Anthropic 最初通过“Project Glasswing”项目提供 Mythos 访问权限，该项目允许经过审核的组织使用该模型测试其代码库，但此后已通过其他受控项目扩大了访问范围。这种安排非常合理，因为如果落入坏人之手，Mythos 可能帮助攻击者利用软件漏洞。Cisco 这笔交易让最终用户接触到 Mythos 的功能，但不允许他们直接使用该模型本身。
Natalie Mead, Anthropic’s head of applied AI for the Asia-Pacific, said the company has a long queue of buyers who want to use Mythos, and working with Cisco will make it easier to satisfy that demand. Anthropic initially made Mythos available under "Project Glasswing," a program that gave vetted organizations access to the model to test their codebases, but has since expanded access through other controlled programs, an arrangement that makes considerable sense because in the wrong hands Mythos could help attackers exploit software flaws. The Cisco deal exposes the functionality of Mythos to end-users, without letting them use the model itself.
记者询问 Mead Anthropic 是否与其他第三方有类似协议，她拒绝置评。这可能是因为在 Cisco 告知我们其使用 Mythos 的情况几小时后，IBM 也宣布将把 Anthropic 的模型集成到其产品中——即 Concert 工具，该工具利用 AI 扫描应用程序的风险和弱点。
The asked Mead if Anthropic has similar deals with other third parties and she demurred. That may be because a few hours after Cisco told us about its use of Mythos, IBM announced that it will also integrate Anthropic’s model into a product – namely its Concert tool, which uses AI to scan applications for risks and weaknesses.
IBM 表示，将使用 Mythos “发现传统扫描器遗漏的弱点，判断哪些构成真实风险，并为开发者批准生成修复方案”。我们已向 Anthropic 询问这些集成是否意味着决定以更多方式向更多用户开放 Mythos。如果收到实质性回复，我们将更新本报道。
IBM says it will use Mythos “to uncover weaknesses traditional scanners miss, determine which pose real risk and generate fixes for developer approval.”We have asked Anthropic if these integrations represent a decision to open Mythos to more users in more ways. We will update this story if we receive a substantial response.
同样在周一，Anthropic 推出了 Mythos 和 Fable 的 5.2 版本，公司称其生成输出的速度比 5.1 版本快多达 30%，且成本更低：每百万输入 token 8 美元，每百万输出 token 40 美元。用户可选择以 “快速模式” 运行这两个模型，这能将 token 生成速度提升多达 2.5 倍，价格翻倍至每百万输入 token 16 美元、每百万输出 token 80 美元。
Also on Monday, Anthropic introduced version 5.2 of Mythos and Fable, which the company says generate outputs up to 30 percent faster than version 5.1 yet cost less: $8 per million input tokens and $40 per million output tokens. Users can choose to run both models in “fast mode,” which speeds up token generation by up to 2.5 times, and doubles prices to $16 per million input tokens and $80 per million output tokens.
Anthropic 称，“对 Fable 5.2 和 Mythos 5.2 的底层模型进行了广泛的发布前安全测试。这包括开发新颖的评估方法，以更深入地洞察沙箱逃逸等特定令人担忧的行为。” 经过这些测试后，公司得出结论：“错位带来的高风险仍然较低，且该模型在大多数对齐指标上的表现与我们近期的模型持平或更好。Fable 5.2 的网络和生物安全防护与 Fable 5.1 一样稳健，但产生的不必要拦截更少，我们还实施了有史以来最强的蒸馏攻击防护。”
后者是一个耐人寻味的承认，因为 Anthropic 认为中国 AI 实验室利用蒸馏技术，在不进行昂贵训练工作的情况下有效克隆其模型。®
Anthropic says it “ran extensive prerelease safety testing on the model that underlies Fable 5.2 and Mythos 5.2. This included developing novel evaluations that give deeper insight into specific concerning behaviors like sandbox escapes.”After those tests, the company concluded “high-stakes risks from misalignment remain low, and the model performed as well as, or better than, our recent models on most alignment measures. “Fable 5.2’s cyber and biology safeguards are as robust as those of Fable 5.1 but produce fewer unnecessary blocks, and we have implemented our strongest-ever protections against distillation attacks.” The latter is a telling admission, given Anthropic believes Chinese AI labs use distillation to effectively clone its models without doing expensive training work. ®