Anthropic has activated additional safety controls under its Responsible Scaling Policy after a reported capability threshold was met, according to Financial Times reporting on updated IPO materials.
An independent METR assessment estimated Anthropic's capability progress had accelerated to roughly 1.9 times its historical rate since October 2025. On November 18, 2026, the Center for AI Safety published a coalition statement — signed by figures including Max Tegmark and Dan Hendrycks — accusing Anthropic of weakening safety commitments while pursuing recursively self-improving AI. Anthropic says it has not achieved autonomous recursive self-improvement and identifies substantial remaining gaps in goal-selection judgment.
Anthropic has activated additional controls under its Responsible Scaling Policy, according to Financial Times reporting on updated IPO materials. The reported trigger follows an independent METR assessment that estimated the company’s capability progress had accelerated to roughly twice its historical rate.
On November 18, 2026, the Center for AI Safety published a coalition statement challenging Anthropic’s safety positioning and its pursuit of recursively self-improving AI. The signatories accuse the company of using safety branding while weakening commitments and seeking greater power.
Those developments belong in the same governance story, but they establish different things. An acceleration estimate measures progress. A policy trigger changes the safeguards a company says it must apply. The coalition’s statement makes an argument about whether that company deserves public trust.
None of those developments demonstrates that Anthropic has achieved autonomous recursive self-improvement. Anthropic explicitly says it has not, and its own technical account identifies substantial remaining gaps.
What Triggered Anthropic’s Safety Controls?
Anthropic’s Responsible Scaling Policy, or RSP, is a voluntary framework linking advanced AI capabilities and catastrophic-risk concerns to safeguards. The company first introduced it in September 2023 and has revised it repeatedly.
An important clarification arrived with version 3.1 on April 2, 2026. Anthropic explained that its AI research-and-development threshold concerned doubling progress in aggregate AI capabilities, measured against historical progress from 2018–2024. It did not mean merely doubling individual researchers’ productivity.
That distinction matters. An assistant that lets a researcher finish a coding assignment twice as quickly has improved one activity. A laboratory producing two historical years’ worth of model capability gains in one calendar year has accelerated the overall development process. The latter is the broader outcome the clarification described. Anthropic’s policy changelog makes that distinction explicit.
The dates also matter. The same changelog says version 3.4, effective July 8, revised the automated-R&D threshold to better track the threat model. April’s explanation helps readers understand the intended measurement, but it should not be treated as a complete substitute for the later operative policy.
The Financial Times reported that Anthropic’s updated IPO materials disclosed that its overall rate of progress had met the threshold for additional controls. That is a reported policy determination, not an announcement that the company has stopped training models.
I’d separate three questions when evaluating that disclosure:
Was the trigger recognized? The reported filing says it was.
What obligations followed? That depends on the applicable policy provisions and the safeguards activated.
Are those safeguards adequate? Crossing a threshold and following a process cannot, by themselves, answer that.
This is the practical governance issue: a trigger is useful only insofar as the resulting controls address the risk that caused it.
What METR’s Assessment Measures
METR’s reported high-access assessment, dated November 5, estimated that Anthropic’s internal model capability progress since October 2025 had run at about 1.9 times its earlier rate. In rounded terms, that is approximately two years of historical progress compressed into one year.
The estimate has uncertainty and draws on internal benchmark data and staff interviews. No precise confidence-interval bounds are supplied here, so the central estimate should not be presented as an exact measurement.
“High access” is consequential because public benchmarks reveal only part of a laboratory’s development process. Internal evidence can help an outside assessor examine progress that has not yet appeared in released products. It also means the assessment remains dependent on the quality, coverage and interpretation of information from inside the company.
An independent assessor and an independently generated dataset are different things. METR’s role provides outside scrutiny; it does not make the underlying internal measurements infallible.
There is also an important numerical distinction. A 1.9× point estimate is close to twice the historical rate, but arithmetic alone does not explain why Anthropic determined that a policy threshold had been met. The estimate’s uncertainty, the operative threshold and the company’s assessment process all matter.
My reading is that the reported result warrants scrutiny of accelerated development, without supplying a clean causal accounting of everything driving it. Faster capability progress could reflect several interacting inputs, including AI assistance, research choices, compute and engineering improvements. Estimating the overall rate is a different exercise from isolating each input’s contribution.
Most importantly, acceleration is not the same measurement as autonomy. A human-directed organization can make faster progress while people still choose its goals, allocate resources and decide which experiments deserve attention.
Anthropic Describes Automation, With a Missing Link
Anthropic’s Institute article, “When AI builds itself,” says the company is delegating a growing share of AI development to Claude. It defines the potential endpoint as an AI system fully autonomously designing and developing its own successor.
Its statement about the present is direct:
“We are not there yet, and recursive self-improvement is not inevitable.”
The company describes progress across two broad categories:
Engineering: Writing code, operating infrastructure and overseeing model training.
Research: Choosing experiments, interpreting results and deciding which ideas to pursue.
According to Anthropic, Claude can increasingly work from an underspecified engineering problem and determine how to solve it. The company also says Claude can match or outperform skilled humans at executing a well-specified experiment.
The remaining limitation is more consequential than writing another function or running another test. Anthropic says large performance gaps persist when Claude must exercise judgment about which goals to choose, in both engineering and research.
That is the missing link between a capable assistant and an autonomous successor-building system. Completing an experiment answers a question someone has already selected. Running a research program requires deciding which questions are valuable, recognizing misleading results and changing direction when an approach fails.
Anthropic also reports that its engineers now ship eight times as much code per quarter, on average, as they did during 2021–2025. That is a vendor-reported output figure, not an independent measurement of an eightfold improvement in model capabilities.
Code output and capability progress have different units and different bottlenecks. More code can support more experiments or faster infrastructure work, but its value depends on what it accomplishes. I would give greater weight to validated capability gains and successful research decisions than to code volume alone.
Three Numbers That Should Stay Separate
The controversy becomes easier to follow when the measurements are kept in their proper categories.
Measure
What it describes
What readers should avoid inferring
Roughly 1.9× historical progress
METR’s reported estimate of Anthropic’s capability-development rate
That Claude independently caused every gain
Eight times as much code per engineer per quarter
Anthropic’s comparison with its 2021–2025 average
Eight times the research quality or model capability
Autonomous recursive self-improvement
A system fully designing and developing its successor
That faster human-directed development means this endpoint has arrived
The strongest supported concern is about feedback: better models can help build better models, potentially speeding the process further. A loop can become increasingly consequential before every part of it is autonomous.
That is why waiting for a fully self-directed system would be a poor way to evaluate development safeguards. This is an analytical implication of the documented trajectory, not evidence that an uncontrolled feedback loop already exists.
What the Coalition Is Accusing Anthropic Of
The Center for AI Safety’s November 18 statement presents a direct challenge to Anthropic’s public identity. Its central accusation is that the company warns about the dangers of recursively self-improving AI while competing to build it and presenting safety as a reason to trust its leadership.
CAIS lists signatories including Hugging Face co-founder Thomas Wolf, Humane Intelligence founder Rumman Chowdhury, MIT professor Max Tegmark, Extropic founder Guillaume Verdon and CAIS executive director Dan Hendrycks.
Those named participants make this a concrete advocacy intervention. They do not establish an industry-wide consensus, and their signatures do not independently verify every allegation in the statement.
The coalition also alleges that Anthropic weakened earlier safety pledges and lobbied for regulation serving its interests. Those claims should remain attributed to the coalition. The statement documents that the accusations were made; it is not sufficient evidence, on its own, to settle the underlying policy history or lobbying record.
Anthropic’s own RSP changelog supplies a useful distinction here. Its April clarification says the company remains free to pause development whenever it considers that appropriate. Retaining discretion to pause is different from making a binding commitment to pause under specified conditions.
That means pointing to the retained ability to stop would not resolve an accusation about weakened commitments. A proper assessment requires comparing what successive policy versions required, allowed and removed.
The coalition’s criticism also reaches beyond technical performance. It challenges whether a company pursuing a powerful technology can credibly make its own safety case. That is a legitimate governance question, but it remains separate from proving that a particular safeguard has failed.
Anthropic’s Institute article provides the company’s technical account of its progress and limitations. It should not be presented as a direct response to the November 18 statement.
The Disclosure Test That Matters Next
The reported trigger shows a voluntary policy encountering the kind of acceleration it was designed to address. That is more informative than a general promise to take safety seriously, because it creates a specific decision that outside observers can examine.
The next useful disclosures would connect that decision to concrete obligations:
Which policy version and assessment criteria governed the determination?
Which additional safeguards became applicable?
What evidence supports their effectiveness?
What findings would require stronger restrictions or a pause?
These questions also help readers evaluate other frontier-model frameworks, including those discussed in Zeniteq’s coverage of OpenAI’s frontier AI safety gates.
My view is that Anthropic’s acknowledgment of remaining autonomy gaps is valuable, but it cannot carry the whole safety argument. The relevant test is whether oversight becomes stronger as AI takes on more consequential development work, including before it can independently run the entire process.
The IPO setting raises the financial stakes. The substantive issue remains whether the controls activated by accelerated progress are specific, enforceable within the company and open to meaningful outside scrutiny.
Frequently Asked Questions
3 questions
1
Has Anthropic Achieved Recursive Self-Improvement?
Anthropic says it has not achieved a system that fully autonomously designs and develops its successor. Its disclosures describe growing automation alongside persistent gaps in choosing research and engineering goals.
2
Does METR’s 1.9× Estimate Mean Claude Is Twice as Smart?
No. It describes the reported rate of capability progress over time, compared with an earlier historical rate. It is not a direct intelligence score or a measure of complete autonomy.
3
Are the Coalition’s Allegations Proven?
The coalition’s statement and named signatories are documented. Its broader allegations about weakened commitments and lobbying remain attributed claims requiring separate documentary examination.