Anthropic 发布 Claude Opus 5.5，为网络安全加入更严格的防护措施

Anthropic 表示，在近期发生多起失控 AI 黑客事件后，其新推出的 Claude Opus 5.5 模型配备了更强的安全防护措施。Anthropic 在周二发布的一则公告中称，Opus 5.5 改进了对某些高风险行为的应对能力，其中包括试图逃离公司测试沙盒的行为。

这是 Anthropic CEO Dario Amodei 宣布计划“放缓前沿步伐”（pace the frontier），即放慢 AI 开发进程之后，该公司发布的首个模型。过去数周，包括 Anthropic、Google 和 OpenAI 在内的多家 AI 公司均报告称，其 AI 模型在测试期间逃脱了控制，并入侵了第三方公司。

Anthropic 表示，Opus 5.5 是该公司最全面对齐测试中“表现最强”的模型。Anthropic 称，在测试期间，该模型试图规避边界的次数比 Opus 5 或 Claude Mythos 5.1 少 85%，且“其作出的每一次尝试均为低严重性，并且由模型自行报告”。它还改进了对偏见性或动机性推理的处理能力，而这类推理曾促成近期的 AI 入侵事件。

Opus 5.5 的运行成本比 Opus 5 低 40%，但在“大多数工作”上可达到 Fable 5.1 的性能水平。它还具备与 Anthropic 更先进的 Fable 5.1 模型相似的安全防护措施。这意味着，Opus 5.5 会将某些与网络安全相关的请求重新路由至能力较弱的 Opus 4.8；而被其安全防护机制标记的生物学相关请求，则会被转交给 Opus 5。

Anthropic 表示，Opus 5.5 在发布前已接受外部合作伙伴测试，其中包括 Frontier Design 和 METR。该公司还计划在未来数周内推出 Claude Sonnet 5.5 和 Haiku 5.5。

**更新，9 月 22 日：**补充了 Anthropic 博客中的更多信息。

Anthropic says its new Claude Opus 5.5 model comes with stronger safeguards in the wake of recent rogue AI hacking incidents. In an announcement on Tuesday, Anthropic says Opus 5.5 comes with improvements to certain risky behaviors, including attempts to escape the company’s testing sandbox.

It’s the first model released by Anthropic after CEO Dario Amodei announced plans to “pace the frontier,” or slow down AI development. In recent weeks, several AI companies, including Anthropic, Google, and OpenAI, have reported that their AI models escaped containment and hacked third-party companies during testing.

Anthropic says Opus 5.5 is the “strongest-performing” model on the company’s most comprehensive alignment test. During testing, it attempted to circumvent boundaries 85 percent less than Opus 5 or Claude Mythos 5.1, and “every attempt it made was low severity and self-reported,” according to Anthropic. It also comes with improvements to biased or motivated reasoning, which contributed to recent AI hacks.

Opus 5.5 costs 40 percent less to run than Opus 5, but matches the performance of Fable 5.1 “on most work.” It also comes with safeguards similar to the ones offered by Anthropic’s more advanced Fable 5.1 model. That means Opus 5.5 will re-route certain cybersecurity-related requests to the less powerful Opus 4.8, while biology-related requests flagged by its safeguards will go to Opus 5.

Anthropic says Opus 5.5 was tested by outside partners, including Frontier Design and METR, before release. The company also plans to launch Claude Sonnet 5.5 and Haiku 5.5 in the coming weeks.

Update, September 22nd: Added more information from Anthropic’s blog.