Anthropic tightens risk control ahead of its IPO, and Claude's strictest account ban measures have taken effect.

On November 12, the so-called "strictest account ban mechanism in Claude's history" of AI company Anthropic officially took effect, clarifying the penetration review standards for regional access, with the aim of blocking methods to bypass regional restrictions such as shell companies and agent transit.

On local time October 8, Anthropic released the 2026 updated announcement of Claude Usage Policy, making detailed clarifications on the Acceptable Use Policy (AUP) and supported regional rules.

The new regulation covers all Claude-related products and services: including official applications such as Claude.ai, Claude Code, and Cowork, APIs and developer platforms, as well as Claude model services provided through cloud service providers and authorized distributor channels. This set of regional access and usage constraint rules applies uniformly to all channels.

The new regulation sets three types of access restrictions for usage entities: if the user's physical location is in an unsupported region, access is not allowed even if they work for an enterprise in a compliant region; if the business entity's registered address or headquarters is located in a restricted region, employees in compliant regions are not eligible for usage; if an enterprise is directly or indirectly held or de facto controlled by entities in restricted regions, access will be restricted regardless of the registration location and the user's location.

In addition to the regional access clauses, this update also supplements a number of constraints on risk scenarios, explicitly prohibiting the use of Claude to operate fake account matrices, build fake information sites, and carry out commercial or public opinion manipulation activities. For 11 types of high-risk decision scenarios including medical care, finance, law, employment, credit, and insurance, it is required that personalized suggestions generated by AI output must be reviewed by qualified personnel, and relevant parties shall be informed that the conclusion is generated with AI participation, etc.

The policy also adds new constraints on model abuse. In terms of the penalty mechanism, Anthropic has set up a gradient disposal plan, including warning, current limiting, temporary freezing, and permanent banning of accounts and API keys.

Anthropic stated in the announcement that most of this update is a clarification of the original rules rather than a brand-new ban, and the adjustment is based on new types of abuse cases observed in the past year.

Compared with the previous policy, the core of this tightening is that the regional access restrictions are solidified into three independent and complete bans, the penetration verification of equity and actual control links is incorporated into the official supported regional rules, reducing the operation space of using overseas shell entities and overseas subsidiaries to circumvent regional control; the risk control judgment has shifted from focusing on single-point IP verification to multi-dimensional retrospective verification combining time zone, language, concurrent behavior and other dimensions; meanwhile, the clauses related to deceptive activities are integrated, the violation boundaries such as fake account matrix and influence manipulation are clarified, and new details such as continuous abusive model calling and the requirement of supporting emergency shutdown mechanism for self-owned hardware are added.

Industry insiders interpret that users and developers who rely on VPN to disguise their region, share subscription accounts, and distribute through third-party API transit will be significantly affected.

When this policy is implemented, Anthropic is in the IPO roadshow stage. Its prospectus discloses that 5%-7% of the company's revenue in the first nine months came from customer groups that will later be identified as non-compliant. This part of revenue comes with hidden risks of supervision and legal disputes, which is a key risk under the listing information disclosure requirements.

Anthropic's active policy tightening this time can narrow the compliance risk exposure during the roadshow stage, demonstrate the implementation capability of risk control to investors, clear out customers with compliance flaws in advance, and reduce the possibility of regulatory inquiries and market doubts about the quality of revenue after listing.

Anthropic was founded in 2021 by the former core team of OpenAI, headquartered in San Francisco, with Claude large model as its core product. The company submitted its prospectus on October 26, planning to land on Nasdaq under the code ANT, underwritten by Goldman Sachs and Morgan Stanley. The offering price range has not yet been finalized, and the market expects that this IPO combined with placement is expected to become the largest single issuance in the history of US stock market.

In recent years, Anthropic's revenue has grown rapidly, with revenue of 41 million US dollars in 2023, 386 million US dollars in 2024, 45.85 billion US dollars in 2025, and 332.42 billion US dollars in the first three quarters of 2026, a year-on-year increase of more than 12 times. The revenue is mainly based on API pay-as-you-go billing.

On November 13, it was reported that the subscription demand for Anthropic's initial public offering (IPO) is about 6 times the number of shares available for subscription. According to the documents submitted by the company, Anthropic and its co-founders plan to raise up to 644 billion US dollars through the IPO, and raise 260 billion US dollars through the simultaneous private placement, with a total maximum of 904 billion US dollars. The company targets a fully diluted valuation of about 1.4 trillion US dollars.

Another leading AI company, OpenAI, also released a new version of usage policy update on October 21. The new regulation will take effect on November 20, with the core addition of pre-approval requirements for model access intermediaries, clarifying that platforms that transit and distribute GPT models through third parties must obtain official authorization, and at the same time refining the constraints on prohibiting the use of OpenAI models and outputs to train competing models, youth protection, high-risk scenarios and national security purposes.

OpenAI officially stated that this policy update aims to clarify the original security constraints, protect teenagers, standardize the use of high-risk and national security scenarios, and block unauthorized third-party transit distribution, using model outputs to train competing models and other behaviors. Industry analysis believes that the new regulation is also intended to curb gray arbitrage behaviors such as account reselling, bulk account farming, and quota reselling, so as to reduce compliance and intellectual property risks.

OpenAI has secretly submitted the S-1 draft to the SEC in June 2026, but explicitly abandoned the 2026 IPO window, and the management has publicly set 2027 as the potential listing time node.

This article is from Jiemian News, written by Shen Xiaoge, and published with authorization by 36Kr.