内容来源：https://aibusiness.com/generative-ai/restrictions-fable-5-mythos-5-lifted-anthropic-launches-sonnet-5

内容总结：

AI模型安全风波启示录：企业选择生成式AI需兼顾性能与治理

谷歌云特约报道

随着大语言模型能力持续突破，企业如何选择首个生成式AI应用场景成为关键。行业专家指出，应优先关注能优化人类信息获取体验的领域，同时需建立跨模型架构思维，并将治理机制纳入模型选型决策。

Anthropic三款模型命运迥异
本周，Anthropic同时推进多款模型更新：此前因国家安全原因被美国监管机构封禁的Fable 5和Mythos模型重新上架，同时发布新一代Sonnet 5模型。值得关注的是，亚马逊研究人员曾发现可绕过Fable 5防护机制的方法，促使政府于6月12日以安全为由叫停该模型。Anthropic表示，为尽快恢复服务，团队训练了针对越狱攻击的新型安全分类器，但Mythos 5截至6月30日仍仅限特定用户访问。

Sonnet 5主打高性价比自主能力
不同于前两者的审查风波，Anthropic宣称Sonnet 5在自主运行时可调用浏览器、终端等工具，性能接近Opus 4.8但成本更低。该模型推出限时优惠：即日起至8月31日，输入/输出价格分别为每百万token2美元和10美元，此后恢复至3美元/15美元。

企业应从个案审视系统性风险
RPA2AI Research创始人Kashyap Kompella指出，此次事件为企业敲响警钟：“必须为模型最优性设计架构，避免将关键工作流绑定单一供应商或模型家族。不同任务需匹配不同等级的机器智能——顶级任务用顶级模型，常规任务则可选用成本更低、稳定性更高的替代方案。”

他进一步强调，模型治理范畴已远超防幻觉、防数据泄漏等传统议题，扩展至涉及地缘政治风险的许可证合规问题。企业需重新审视AI战略设计：哪些人员可访问哪些模型？是否存在国别限制？离岸团队能否使用与本土团队相同的工具？若政府指令中断服务该如何应对？

政策与商业的平衡难题
Kompella补充道，OpenAI的GPT-5.6因政府审查延迟发布，印证此类监管非个例。但美国政府同时面临两难：既要防范网络滥用和先进模型能力带来的安全风险，又要维持美国AI产业的全球竞争力。“过度渲染模型危险性可能招致更严格的监管审查，”Kompella警告，“当企业反复向政府强调系统能力超群、难以管控时，监管机构终将当真。”

中文翻译：

由谷歌云赞助
选择您的首个生成式AI应用场景
要开始使用生成式AI，首先聚焦于能够改善人类与信息交互体验的领域。

强大模型的发布表明，企业需要开放接纳不同的AI系统，并将治理机制纳入模型选择考量之中。

在其Fable 5与Mythos模型重新向客户开放的同一天，Anthropic发布了Claude Sonnet 5，称其为迄今为止最具智能体能力的Sonnet模型。

此前亚马逊研究人员发现一种可绕过Fable 5安全防护的方法并发布报告后，美国政府于6月12日以国家安全为由关闭了这两款模型，现已解除相关限制。

Anthropic表示，为尽快恢复Fable 5模型，他们训练了改进型安全分类器，专门针对并拦截曾导致政府实施出口管制的越狱行为。尽管Mythos 5已于6月30日恢复，目前仍仅限部分用户使用。

而关于Sonnet 5，Anthropic称这款新模型可在自主运行时使用浏览器和终端等工具，其运行水平此前通常需要更昂贵的模型才能达到。Anthropic表示该模型性能接近Opus 4.8，但价格更低。

即日起至8月31日，该模型定价为每百万输入token 2美元、每百万输出token 10美元；此后价格将调整为每百万输入token 3美元、每百万输出token 15美元。

尽管Anthropic似乎已与美国政府解决此障碍，但企业仍应借此事件评估当前使用的模型。

RPA2AI Research公司首席执行官兼创始人Kashyap Kompella表示："对企业而言，这深刻表明他们应当为模型最优性设计架构。"他补充道，企业不应将关键工作流锁定在单一供应商或模型系列中。

"企业需要将不同应用场景匹配到不同级别的机器智能，"Kompella说，"某些任务需要最顶尖的模型，而许多其他任务则可运行在性能较低、更便宜、更稳定或全球可及性更高的模型上。"

Anthropic事件也向企业释放信号：模型治理已不仅限于防范幻觉、数据泄露、员工滥用和AI偏见，更扩展至模型许可、地缘政治及监管可用性风险。

"对企业来说，这改变了AI策略的设计方式，"Kompella表示，"企业需要追问：组织内哪些人员可访问哪些模型？是否存在国家特定限制？某些模型是否仅面向审核客户或仅限特定场景使用？离岸团队是否能与美国团队使用相同工具？若政府指令中断访问会怎样？"

他补充道，OpenAI的GPT-5.6因政府审查而延迟全面发布，也表明Anthropic事件并非孤例。

尽管如此，美国政府也需平衡对模型的看法。Kompella指出，随着政府目标超越中国、欧洲及其他开源与主权AI项目，"美国希望管控风险，但不愿将全球用户推离美国AI基础设施。"

于是困境浮现：政府需在对网络滥用和高级模型能力的国家安全关切、保持竞争力与维护美国AI模型供应商商业利益之间寻找平衡。

Fable 5与Mythos 5事件可能促使Anthropic谨慎推广其模型。随着近期Mythos和Project Glasswing的发布，Anthropic称其模型为最强大，甚至暗示提供公共访问可能带来危险。

"但安全论述达到某个临界点便会招致监管介入，"Kompella说，"若一家公司反复告知政府其系统异常强大、难以控制且存在潜在危险，政府终将认真对待这种说法。"

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
The release of the powerful models shows that enterprises need to be open to different AI systems and consider governance as part of choosing models.
On the same day it made its Fable 5 and Mythos models available to customers again, Anthropic released Claude Sonnet 5 as the most agentic Sonnet model yet.
The U.S. government lifted restrictions on the two models after it shut them down on June 12 due to national security concerns, following a report by Amazon researchers who found a method to bypass Fable 5’s safeguards.
Anthropic said to get Fable 5 restored as quickly as possible, it trained an improved safety classifier that targets and blocks the jailbreaking behavior that led the government to apply export control to the models. While Mythos 5 has been restored as of June 30, it is still only available to a limited number of users.
In the case of Sonnet 5, however, Anthropic said the new model can use tools such as browsers and terminals while running autonomously at a level that would otherwise require more expensive models. The model’s performance is close to that of the Opus 4.8 model, but at a lower price, Anthropic said.
The model is available at $2 per million input tokens and $10 per million output tokens from now till August 31. After that, pricing moves to $3 per million input tokens and $15 per million output tokens.
While it appears that Anthropic has cleared this hurdle with the U.S. government, enterprises should still use this episode to evaluate the models they currently use.
“For enterprises, this drives home that they should architect for model optimality,” said Kashyap Kompella, CEO and founder of RPA2AI Research.
He added that enterprises should not lock critical workflows into a single vendor or model family.
“Companies will need to match use cases to different grades of machine intelligence,” Kompella said. “Some tasks need the very best model available. Many others can run on a less capable, cheaper, more stable or more globally available model.”
The situation with Anthropic should also signal to enterprises that model governance is more than just looking out for hallucinations, data leakage, employee usage and AI bias. It has expanded to include model licensing and geopolitical and regulatory availability risk.
“For enterprises, this changes how AI strategy should be designed,” Kompella said. “Companies will need to ask: Who in the organization can access which models? Are there country-specific limitations? Are some models available only to vetted customers or only for certain use cases? Can offshore teams use the same tools as U.S. teams? What happens if a government directive interrupts access?”
He added that the delay in the full release of OpenAI’s GPT-5.6, while the government reviews it, also shows that the Anthropic incident is not isolated.
Nevertheless, the U.S. government also must balance its views on models. With the government aiming to top China, Europe and other open source and sovereign AI programs, “the U.S. wants to manage risk, but it does not want to push the world away from American AI infrastructure,” Kompella said.
So, a conundrum arises: the government needs to manage national security concerns about cyber misuse and advanced model capabilities while remaining competitive and safeguarding the commercial interests of American AI model providers.
The Fable 5 and Mythos 5 situation will likely lead Anthropic to be careful about how it markets its models. With the recent release of Mythos and Project Glasswing, Anthropic has called its models the most powerful and even made it appear as if providing public access would be dangerous.
“But there is a point at which safety rhetoric becomes a regulatory invitation,” Kompella said. “If a company repeatedly tells governments that its systems are exceptionally capable, difficult to contain and potentially dangerous, governments may eventually take that claim seriously.”