Anthropic’s S-1 Filing Flags Risk of AI Self-Improvement
In a revised SEC filing, Anthropic says internal estimates for future models suggest it may imminently trigger the AI-acceleration threshold set by its Responsible Scaling Policy.
Anthropic's revised S-1/A, filed with the SEC on November 2, 2026 ahead of its Nasdaq listing under ticker ANT, discloses that the company's internal estimates for future models — not its currently shipping ones — suggest it "may imminently" meet the threshold its own Responsible Scaling Policy (RSP) sets for "dramatically accelerated progress attributable to AI." That threshold is the formal marker Anthropic uses to track progress toward what the filing calls recursive self-improvement: the point at which AI systems become capable of meaningfully designing and building their own successors. Anthropic says no current model has crossed the line, that it is still assessing the underlying data, and that it is weighing changes to the RSP specifically to keep humans in control of frontier model development.
- Filed: November 2, 2026, as Amendment No. 1 to Anthropic's Form S-1 (Registration No. 333-299633).
- The Risk Factors section says using AI to build more capable AI "could also pose a considerable safety risk," an area covered by the Responsible Scaling Policy as Anthropic approaches "what has been termed recursive self-improvement."
- Anthropic says it has historically not identified a significant acceleration in its development rate "until after it has occurred," limiting its ability to forecast the pace of change.
- Internal estimates and expectations for future models — not current ones — indicate Anthropic "may imminently" meet the RSP's dramatically-accelerated-progress threshold.
- No current Anthropic model has been designated as exceeding that threshold; the company says it continues to assess the data.
- Anthropic says it does not expect to be able to safely let Claude fully autonomously design and build successive frontier models without meaningful human supervision, and is considering RSP updates to keep humans in control.
What Anthropic's S-1/A actually says
The disclosure sits in two places in the amended S-1: the Risk Factors section and the Business section. In Risk Factors, Anthropic describes its reliance on using its own models to accelerate research — what it calls "closing the loop," where "successive versions of our models assist in building the next, with the goal of compounding capability gains over time." The filing warns that this reliance cuts both ways: it is central to staying competitive, but it also means "our increasing use of AI systems to accelerate our own research and development may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures."
The filing is explicit that Anthropic's visibility into this acceleration is limited. "Much of our evidence regarding the pace of such AI-driven acceleration is derived from lagging indicators," it states, "and our ability to forecast or measure the rate at which our capabilities are advancing… is limited. As a result, we have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred." That sentence is immediately followed by the core disclosure: "our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling Policy's definition of dramatically accelerated progress attributable to AI. We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data."
In the Business section, Anthropic puts the stakes more plainly: "Using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress. This is one of the areas covered by our Responsible Scaling Policy, which sets out our plans to monitor and mitigate the risks as we approach what has been termed recursive self-improvement." The company adds that it "do[es] not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible," and says it is "considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic." For background on the broader filing, Pandromeda has previously covered the S-1 filing's ticker and valuation details and the original S-1 filing announcement.
What "recursive self-improvement" means here
Recursive self-improvement (RSI) is a long-standing concept in AI safety research describing a feedback loop in which an AI system becomes capable of improving its own design, training process, or successors, with each iteration making the system more capable of driving the next one — potentially without a ceiling set by human researchers. Wikipedia's entry on the concept traces the idea back decades in AI theory, but Anthropic's S-1/A is one of the first instances of a frontier AI lab naming RSI directly in a securities filing's risk disclosures, rather than only in research essays or safety blog posts.
Anthropic is not claiming RSI has happened. The filing's language is carefully hedged: current models have not crossed the defined threshold, and the company is still assessing data for what might be imminent in future models. The disclosure is forward-looking risk language required in a securities filing, flagging a scenario investors and regulators should understand as plausible rather than confirmed.
The Responsible Scaling Policy threshold, explained
Anthropic's Responsible Scaling Policy is the company's own framework for defining AI capability thresholds and the safeguards that must be in place before Anthropic will train or deploy a model that crosses them. The policy has been revised multiple times as Anthropic's models have advanced, and the specific threshold referenced in the S-1/A — dramatically accelerated progress attributable to AI — has its own revision history inside the RSP itself.
| RSP update | What changed on the AI R&D threshold |
|---|---|
| v2.1 (March 31, 2025) | Split the AI R&D threshold into two tiers: the ability to cause dramatic acceleration in the rate of effective scaling, alongside full automation of entry-level AI research work. |
| February 10, 2026 | Required that once a model crosses the "AI R&D-4" capability threshold, Anthropic must develop an affirmative misalignment-risk case before proceeding. Anthropic states Claude Opus 4.6 does not cross this threshold. |
| v3.1 (April 2, 2026) | Sharpened the definition of dramatically accelerated progress, illustrated by the example of AI compressing roughly two years of 2018–2024-era AI progress into a single year. |
| v3.4 (July 8, 2026) | Revised the automated-R&D threshold again; Anthropic's published policy does not detail the specific change. |
This is the same threshold language the S-1/A invokes when it says future models "may imminently" meet the definition. In other words, the filing is not describing a brand-new concept — it is reporting that Anthropic's own internal forecasts for upcoming models are now running up against a bar the company set for itself roughly 18 months earlier.
No current model has crossed the line, Anthropic says
The filing is careful to separate the forward-looking estimate from the present state of its deployed models. Anthropic states plainly that it "has not designated any of our current models as exceeding this threshold," and that it "continue[s] to assess the underlying data." That distinction matters: the risk being disclosed is about the trajectory implied by internal projections for future model generations, not a claim that an existing Claude model is already self-improving or has been secretly operating past a safety line. Anthropic also says it is "evaluating the implications of meeting the AI research and development capability threshold" and will "release further reports in the future" under the RSP as that evaluation continues, per the S-1/A text.
Why Anthropic says it can't see acceleration coming
A recurring theme in the filing is Anthropic's acknowledgment that it has limited ability to measure its own acceleration in real time. The company says its evidence is "derived from lagging indicators," meaning the data confirming an acceleration happened typically only becomes clear after the fact. That is why the filing frames its past experience the way it does: Anthropic says it has "in the past not identified… a significant acceleration" in its development pace "until after it has occurred." The S-1/A ties this directly to the heavy internal use of Claude across Anthropic's own research and engineering pipeline. Elsewhere in the Business section, Anthropic discloses that over 90% of the contributions to its codebase in September 2026 were written by Claude, and that the average daily volume of code contributed per engineer in the three months ended September 30, 2026 was over 45 times the average daily volume in the year ended December 31, 2024 — concrete evidence, in the company's own telling, of how far "closing the loop" has already progressed on the engineering side even as the broader AI R&D threshold has not been crossed.
Context: Anthropic's June 2026 essay on RSI
The S-1/A disclosure builds on ground Anthropic laid in June 2026, when it published "When AI builds itself," a research essay from its Anthropic Institute exploring how AI-assisted AI development might play out. That essay, published by Jack Clark and Marina Favaro according to Anthropic's own site, laid out three scenarios for how the trajectory could unfold: a stalled plateau in which progress levels off; compounding efficiency gains in which AI automates much of development while humans still set direction and judge results (which Anthropic called the most likely path); and full recursive self-improvement, in which AI systems design and refine their own successors largely unconstrained by human researcher throughput. The essay also disclosed that, as of May 2026, more than 80% of the code merged into Anthropic's codebase was authored by Claude — a figure the S-1/A's September 2026 update (over 90%) suggests has kept climbing. The essay stated outright that full RSI "is not current reality" and is "not inevitable," framing the June piece as an early warning rather than an announcement. The S-1/A is the first time this same risk has been formalized as a disclosure in a securities filing rather than a standalone research post.
Keeping humans in control: policy changes ahead
The clearest statement of intent in the filing is Anthropic's plan to revisit its own safety policy in response to this trajectory. The company says it does "not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible." Because of that, Anthropic says it is "considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development." The filing does not specify what those updates would look like, but it links the commitment to the same precautionary mechanisms Anthropic says it has already used — including, the filing notes as an example, delaying the release of models and limiting their initial availability, citing its handling of the model Anthropic refers to internally as Mythos under what the filing calls Project Glasswing. That example refers to Claude Mythos, a restricted Anthropic model that has been kept under tighter access controls than its general-release Claude models.
Why this matters for the IPO
Risk Factors sections exist precisely so that a company going public discloses material risks to prospective investors, and Anthropic's framing here is unusually direct for the AI industry: it is telling investors, in a legal filing, that the pace of its own AI-driven progress could outrun its ability to monitor or control it, and that this is a plausible near-term scenario for its next generation of models rather than a distant hypothetical. For a company asking public markets to value it on the strength of continued frontier-model leadership, acknowledging that leadership itself could become harder to safely manage is a notable disclosure. It also gives regulators, AI-safety researchers and competing labs a documented, SEC-filed reference point for how seriously Anthropic is treating RSI as a near-term governance problem rather than only a research topic.
What to watch next
Three things will show whether this disclosure was a one-off caution or the start of a pattern. First, whether Anthropic publishes the "further reports" on the AI R&D capability threshold that the filing promises under the RSP, and what those reports say about where upcoming models land relative to the threshold. Second, whether Anthropic follows through on the RSP updates it says it is "considering" to keep humans in control of frontier model development, and what concrete mechanisms — additional sign-off requirements, slower release cadences, new evaluation gates — those updates include. Third, how the disclosure is treated during the IPO process itself: whether research analysts, regulators or rival labs cite the S-1/A's RSI language as a new benchmark for how AI companies should talk about self-improvement risk in public filings. Anthropic's own position, stated in both the June essay and the November S-1/A, is that it is not there yet — but that it is close enough to be disclosing the possibility to investors rather than waiting for it to show up in a future filing.
Frequently asked questions
What did Anthropic's S-1/A say about AI self-improvement?
In its revised S-1/A filed with the SEC on November 2, 2026, Anthropic said its internal estimates and expectations for future models, rather than its current ones, suggest it may imminently meet its Responsible Scaling Policy's defined threshold for dramatically accelerated progress attributable to AI — the marker it uses to track proximity to recursive self-improvement.
Has any current Anthropic model achieved recursive self-improvement?
No. Anthropic states in the filing that it has not designated any of its current models as exceeding the relevant Responsible Scaling Policy threshold, and that it continues to assess the underlying data.
What is recursive self-improvement (RSI)?
RSI describes an AI system becoming capable of meaningfully improving its own design, training process, or successor models, potentially compounding capability gains with less reliance on human researchers at each step.
What is Anthropic's Responsible Scaling Policy?
It is Anthropic's own framework defining AI capability thresholds and the safeguards that must be in place before training or deploying a model that crosses them, including a specific threshold for dramatically accelerated, AI-attributable progress.
Is this related to Anthropic's June 2026 essay on AI building itself?
Yes. Anthropic's June 2026 essay, "When AI builds itself," explored scenarios ranging from a stalled plateau to full recursive self-improvement and said the concept was not yet current reality. The November 2026 S-1/A is the first time Anthropic has formalized this same risk as a disclosure in a securities filing.
What is Anthropic doing in response to this risk?
Anthropic says it does not anticipate safely letting Claude fully autonomously design and build successive frontier models without meaningful human supervision, and that it is considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development.
Sources
- Anthropic, PBC — Form S-1/A (Amendment No. 1), filed Nov. 2, 2026sec.gov
- Anthropic's Responsible Scaling Policyanthropic.com
- Anthropic: "When AI builds itself"anthropic.com
- Wikipedia: Recursive self-improvementen.wikipedia.org
Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.
