Anthropic has announced a transition from covert to overt model downgrades for users engaged in specific high-risk research areas after facing criticism from the developer community.

在面臨開發者社群的批評後，Anthropic 宣布將針對從事特定高風險研究領域用戶的模型降級方式，從隱蔽轉為公開。

The controversy centers on Fable 5, a restricted version of the Mythos model developed under Project Glasswing—a consortium including Apple, Google, and Microsoft aimed at securing internet infrastructure. While Mythos remains restricted to prevent the exploitation of zero-day vulnerabilities, Fable 5 was released to the public with embedded safety classifiers. These classifiers were designed to automatically downgrade user requests to the less capable Opus 4.8 model when the system detected research pertaining to frontier-level large language models (LLMs) or specialized chip architecture. Initially, this transition occurred without user notification, a practice documented in a 319-page system card but omitted from the user interface.

此次爭議的核心在於 Fable 5，它是 Project Glasswing（一個由 Apple、Google 和 Microsoft 組成，旨在保障網路基礎設施的財團）開發的 Mythos 模型的限制版本。雖然 Mythos 仍受到限制以防止 0-day 漏洞被利用，但 Fable 5 在對外發佈時內建了安全分類器。這些分類器的設計旨在於系統偵測到研究涉及前沿大語言模型 (LLM) 或專門晶片架構時，自動將用戶請求降級至能力較低的 Opus 4.8 模型。起初，此轉接過程是在未通知用戶的情況下進行的，雖然在一份 319 頁的系統卡中有所記錄，但在用戶界面中被省略了。

Stakeholder reactions have been bifurcated. Cybersecurity experts, including representatives from the SANS Institute, contend that these restrictions impede legitimate defensive research and the development of next-generation forensic tooling. Conversely, other analysts suggest that the restraint exercised in the model's release is a necessary precaution against the proliferation of dual-use capabilities. Furthermore, the company's data retention policy—mandating a 30-day window for Fable and Mythos to facilitate safety classification—has prompted legal scrutiny from corporate partners such as Microsoft, who typically require zero-data-retention agreements.

利益相關者的反應分成了兩派。包括 SANS Institute 代表在內的網路安全專家認為，這些限制阻礙了合法的防禦性研究以及下一代鑑識工具的開發。相反，其他分析師則認為，模型發佈時所採取的克制是防止雙用途能力擴散的必要預防措施。此外，該公司的數據保留政策——規定 Fable 和 Mythos 需有 30 天的窗口期以利於安全分類——引起了如 Microsoft 等企業合作夥伴的法律質詢，因為後者通常要求零數據保留協議。

Anthropic has characterized these safeguards as essential for maintaining the strategic technological advantage of the United States and its allies against foreign adversaries. However, external observers, including academic faculty from Simon Fraser University, posit that these measures also serve a commercial function. By restricting the development of competing AI systems and preventing 'distillation'—the process by which rivals utilize a superior model's outputs to train their own—Anthropic may be attempting to mitigate the market pressure exerted by lower-cost, open-weight models from entities such as Xiaomi and Z.ai.

Anthropic 將這些防護措施描述為維持美國及其盟友面對外國對手時，擁有戰略技術優勢的必要手段。然而，包括 Simon Fraser University 教職員在內的外部觀察者認為，這些措施同樣具有商業功能。透過限制競爭 AI 系統的開發並防止「蒸餾」（即競爭對手利用更強模型的輸出來訓練自有模型的過程），Anthropic 可能試圖減輕來自小米 (Xiaomi) 和 Z.ai 等實體推出之低成本、開源權重模型的市場壓力。

Anthropic has apologized for the lack of transparency and has implemented a visible fallback mechanism, though the underlying restrictions on frontier AI development remain in effect.

Anthropic 已為缺乏透明度道歉，並實施了可見的後備機制，儘管針對前沿 AI 開發的底層限制依然有效。