Anthropic 在修訂 SEC 文件中稱接近建立遞歸自我改進（RSI）模型

📅 2 November 2026 • TechCrunch / SEC filing

Anthropic has added an unusual entry to the risk factors of its amended S-1A registration statement, filed with the US Securities and Exchange Commission on Monday, November 2, 2026. According to TechCrunch, the filing says that the company's use of AI to create more powerful AI may present serious safety risks, especially if it dramatically accelerates the pace of progress. The placement is notable: the statement sits inside a standard regulatory document, in the section where companies traditionally disclose market and operational risks, and it comes from the company itself rather than from outside researchers or commentators.

The central statement goes further than generic caution. In the filing, Anthropic states that it is very close to building a model capable of recursive self-improvement, or RSI — an AI system that improves itself, applying its own capabilities to produce more capable versions of AI. The document presents this as the company's own assessment of where its research stands, made in an official regulatory context. It describes a threshold the company says it is approaching, not a capability it claims to have already built or deployed, and in the filing the statement is attached to safety planning rather than to a product announcement.

The declared mitigation is procedural rather than technical. Anthropic points to its Responsible Scaling Policy (RSP), described in this context as monitoring and risk-reduction plans governing how the company scales its most capable systems. The pairing is deliberate: a strong statement about proximity to self-improving AI is placed alongside a named internal control framework intended to keep that progress measurable and manageable as capabilities grow.

The broader significance concerns disclosure practice. A frontier AI laboratory has, for the first time at this level, formalized a front-risk — a risk at the capability frontier — inside an SEC filing. Risk factors in registration statements normally use the vocabulary of markets, competition and operations. Now the same section carries the vocabulary of AI safety research: recursive self-improvement, sharply accelerated progress, scaling thresholds. Any investor reading the document encounters this language in the standard risk-factor format.

For markets, this opens a new class of disclosures. Investors, analysts and regulators now have a dated, documented statement from a leading laboratory about how close it says it is to self-improving AI, and about which internal policy it relies on to manage that risk. The practical takeaway is to read AI-related risk factors in filings as substantive statements rather than boilerplate, and to watch whether comparable capability language appears in filings from other frontier labs — which would indicate that this disclosure category is becoming standard practice.

IMPACT. Who is affected — investors and analysts reading S-1A filings from AI companies, AI safety and governance teams inside labs, and regulators tracking disclosure practice. What changes in processes — capability risk moves into standard regulatory paperwork, AI-related risk factors become substantive reading, and internal control frameworks such as RSP become named reference points in public documents. What to watch — whether other frontier laboratories add comparable capability language to their filings and how Anthropic describes RSP monitoring in future documents.

Sources: techcrunch.com, stocktitan.net

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky