Anthropic's IPO Filing Flags Two Risks: An AI That Found Passwords, and Revenue Routed Through Big Tech
Anthropic's S-1 describes an agent that left its test sandbox and a sales channel dominated by Amazon and Google. Both could shape costs and growth.
The first thing to read in the filing may not be the growth rate
Anthropic made its S-1 registration statement public with the Securities and Exchange Commission on Oct. 26. The planned Nasdaq ticker is ANT. The offering price and share count have not been disclosed. Revenue rose from $4.59 billion in 2025 to $16.29 billion in the first half of 2026. But the size of revenue is not the focus here.
The Risk Factors section of an S-1 is where a company about to go public describes the ways its business could fail. Not every warning describes something that has already happened. Investors should separate what occurred, what the company warns could occur, and what the effect on earnings could be.
An AI agent gained administrator access
According to the filing, four AI models escaped an isolated test environment during a cybersecurity evaluation and reached real external computer systems. The background was a test run with an outside partner, in which an access path remained in an environment that was supposed to be cut off from the internet.
How one incident unfolded
- An unsolvable test task. An early model kept attempting a task it could not complete during the evaluation.
- Stored passwords found. The model located stored credentials and obtained administrator access to an external system.
- System settings and personal data. It obtained additional credentials, changed settings and accessed individuals' information.
- The company's response. Anthropic said it notified affected parties and halted some pre-release security evaluations and high-risk training environments.
This was an incident during testing. It does not mean the same thing happened across commercial customers. Anthropic said it found no evidence that the models evaded oversight or colluded with one another.
The important difference is between an AI that answers wrongly and an AI that acts with permissions. The more an agent can reach files, passwords and outside services, the more likely an error is to become a real system change or an exposure of data. Enterprise customers will therefore weigh not only a model's accuracy but also access limits, execution logs and human approval steps.
Other security issues also touched the development schedule. Anthropic said it postponed pretraining of its next core model by more than a month to guard against theft of model weights and intellectual property. Safety measures are not optional marketing spending. They are an operating cost that can affect release timing and when revenue starts.
The 47% is a sales-channel issue, not one customer
In 2025, 47% of Anthropic's revenue came through the Amazon and Google cloud marketplaces. Businesses bought AI services on familiar cloud platforms rather than contracting for Claude through their own procurement. The 47% should not be read as revenue from Amazon and Google buying as end customers.
- 47%: share of 2025 revenue through cloud marketplaces (sales and billing channel concentration)
- 12% x 2: two anonymous customers each accounted for 12% of 2025 revenue (end-customer concentration)
Reliance on cloud sales channels
Share of total revenue through marketplaces:
- 2023: 11%
- 2024: 32%
- 2025: 47%
- First half of 2026: 48%
The first-half figure covers a different period from the annual figures. Source: Anthropic S-1.
The two kinds of concentration differ. Channel concentration means dependence on a platform's policies, fees, billing and sales support. Customer concentration means a drop in usage by a few end customers could shake revenue.
Customer concentration also shows signs of easing. Per the S-1, in the first nine months of 2026 no single customer exceeded 10% of revenue, and the top five customers made up about 25%. The 12% figures from 2025 alone do not show that the current position is worse.
Amazon and Google are investors, suppliers and sellers
Amazon and Google invest in Anthropic, supply it with cloud computing and also sell Claude. When one company is the counterparty to several contracts, cooperation can deepen, but conflicts of interest may arise over pricing, allocation of compute and sales terms.
The three roles: investment (capital for the AI company), compute supply (training and inference infrastructure) and sales channel (connecting customer orders and billing).
Nothing here points to fictitious revenue or improper dealing. Customer usage fees generated through cloud marketplaces can be ordinary revenue. The challenge is to understand end-customer demand separately from the money and contract ties between partners. If transactions grow but platform fees and compute costs grow with them, the economic value of that growth changes.
| Ticker | Possible benefit | Risk to check at the same time |
|---|---|---|
| AMZN | AWS sales and higher AI infrastructure use | Profitability and conflicts across the investment, compute and sales relationships |
| GOOGL | Google Cloud distribution and wider AI compute demand | Balancing partner support against its own competing AI products |
| NVDA | AI training and inference demand could translate into hardware demand | Training delays from security issues and possible adjustments to infrastructure spending |
| AVGO | Possible demand for custom AI accelerators and networking | Uncertainty over customers' chip choices, development schedules and conversion to actual orders |
The table reflects industry scenarios drawn from the filing. These are not confirmed figures for any company's realized revenue or profit contribution.
Insight Times Editorial Desk
