Amodei says Anthropic’s safety delay may have cost it an AI lead
Anthropic first released Mythos-class models in February 2026, but CEO Dario Amodei said the company delayed wider access until June so it could build stronger protections against cyberattacks.
Amodei explained this decision in a long post on X on Tuesday. He was responding to criticism from Nathan Calvin and defending Anthropic's approach to AI safety. He admitted that the delay may have cost Anthropic an advantage over competitors. He argued that releasing the models earlier would have given organizations less time to prepare their computer systems for increasingly powerful AI-driven cyberattacks.
“We likely sacrificed a significant chunk of a several-month model lead,” Amodei writes.
Anthropic told The New Stack that the delay follows a promise in its Responsible Scaling Policy. The policy requires strong evidence that the risk of a catastrophe is under control. It also says the company should delay development and release for as long as necessary, until it no longer believes it has a significant lead over competitors.
Thank you for this, @_NathanCalvin. I appreciate and am a fan of the investigative work you’ve done on behalf of AI safety. I believe your comments come from a constructive and honest place, and that the public benefits from having people whose job is to examine the actions… https://t.co/QZgPvh4tX4
— Dario Amodei (@DarioAmodei) October 21, 2026
Debating a strategy under scrutiny
In his criticism posted Monday, Calvin argues that Anthropic is increasingly acting like a company racing to build superintelligence while weakening its promises to prevent dangerous AI outcomes.
He points to more than $500 billion in computing deals, the company's growing use of AI to develop new models, and changes to its Responsible Scaling Policy. This policy sets safety rules that are meant to apply as AI models become more powerful.
Thank you for taking the time to respond in detail over the weekend. I realize that writing legislation is difficult, and some changes and responses to feedback are both unavoidable and appropriate.
— Nathan Calvin (@_NathanCalvin) October 19, 2026
However, I feel that your response still does not really address the main issue…
“When Anthropic strays and makes mistakes, why do the mistakes systematically tend to guard and favor Anthropic’s interests?”
“When Anthropic strays and makes mistakes, why do the mistakes systematically tend to guard and favor Anthropic’s interests?” Calvin asks.
In February, Anthropic updated its policy to version 3.0. It replaced earlier rules, which Amodei now describes as too strict, with an approach that focuses on being open about decisions, reporting risks, and using independent evaluations. The current version is 3.4, which took effect on July 8.
In its Frontier Safety Roadmap, Anthropic says it will “strive to avoid situations where we revise the goals in a less ambitious direction because we simply can’t execute.”
Why Anthropic removed its safety promise
The February update also removed a promise made in 2023: Anthropic would never train an AI system unless it could guarantee in advance that its safety measures were good enough. Calvin argues that this change weakens the company's promise to slow down or stop development on its own.
Anthropic says the updated policy is its strongest version yet for public accountability and openness.
Anthropic's chief science officer, Jared Kaplan, explained the reasoning in an interview with Time magazine for a story published in February. He said the decision was based on practical concerns.
“We felt that it wouldn’t actually help anyone for us to stop training AI models,” Kaplan says. He adds that making such promises alone did not make sense “if competitors are blazing ahead.”
“Our strategy has always been what I’ve called the ‘race to the top,’”
Amodei makes a similar argument in his Tuesday post on X.
“Our strategy has always been what I’ve called the ‘race to the top,'” he writes. He argues that Anthropic can do more to shape safety standards across the AI industry by staying at the leading edge of technology.
He admits that Anthropic could have handled three of Calvin's criticisms better, including its work on the draft bill. He calls the criticism of the Responsible Scaling Policy “the most substantive item.” However, he argues that the company has repeatedly given up business advantages to improve safety. In addition to delaying Mythos, he points to strict safeguards against biological risks, support for AI regulation, and export controls on advanced chips and chipmaking equipment.
Amodei argues that these decisions show the company is willing to put safety before short-term profits, even when doing so is costly. He says Anthropic's overall strategy has remained consistent.
April–June: Anthropic gives more people access to Mythos
Anthropic first released Claude Mythos Preview in April through Project Glasswing. Selected partners, including Amazon, Apple, Microsoft, CrowdStrike, and the Linux Foundation, received access to the model's ability to find security weaknesses so they could use it for defensive work. Ordinary developers could not access it.
On June 2, Anthropic expanded the program to about 150 more organizations across more than 15 countries. Anthropic warns that a major attack could affect more than 100 million people for most of these partners, based on its own estimate. It also says that “hundreds of thousands of organizations, researchers, and maintainers” will probably need access to advanced cybersecurity tools before it can release Mythos-level models to a much wider audience.
June: Fable is released and then withdrawn
Anthropic released Claude Fable 5 on June 9. This Mythos-class model included extra safeguards against cybersecurity, biological, and chemical risks. The company said these protections made it suitable for general use. Its related model, Mythos 5, used the same underlying model but had fewer cybersecurity restrictions and remained available only to trusted partners. Anthropic explained the difference in its launch announcement.
Three days later, on June 12, a U.S. government export-control order forced Anthropic to suspend both models. The order prohibited foreign nationals, including Anthropic's own employees, from accessing them. Because the restriction applied to foreign nationals everywhere, Anthropic said the only way to follow it was to disable the models for all customers. The company publicly disagreed with the order.
The Commerce Department lifted the restrictions on June 30. In its announcement confirming the decision, Anthropic described plans to work more closely with the government on testing models before release, sharing information, and setting cybersecurity standards for future models. Fable 5 returned worldwide on Anthropic's own platforms on July 1. Mythos 5 had already become available again to selected U.S. organizations on June 26.
Access rules now determine which models developers can use. Anthropic's expanded Cyber Verification Program gives cybersecurity professionals access to its most powerful models, including Claude Mythos 5.1. The program has three levels based on the type and scope of security work people do. Existing Glasswing members can use the level with the fewest restrictions. New organizations applying for that level must go through a review involving Anthropic and the U.S. government.
In Anthropic's own tests, the most restricted level, Defense Access, blocked Claude Opus 5.5 at some point in 46 of 50 trials on CyScenarioBench, a test that measures performance across several stages of offensive cybersecurity work. Red Team Access blocked the model in none of the trials.
The announcement does not include a similar test measuring defensive cybersecurity work. This makes it difficult to tell how often the restrictions might get in the way of legitimate security tasks. Anthropic acknowledges that its safeguards sometimes wrongly block secure coding work and provides a form for users to report these mistakes.
Controlling the pace of AI and planning policy
Amodei acknowledges that AI development may soon move faster than the safeguards designed to control it. “We are nearing that point,” he writes in his post on X.
Anthropic's research helps explain the company's concern. In its June report, “When AI builds itself,” Anthropic says that, as of May, Claude wrote more than 80% of the code that was accepted into the company's codebase.
This figure measures lines of production code credited to Claude. It does not mean that Claude independently controls the company's engineering decisions. Anthropic calls the figure a conservative estimate. Company leaders have separately estimated that Claude writes 90% or more of the code when scripts and experimental code are also counted.
“We are not there yet, and recursive self-improvement is not inevitable.”
“We are not there yet, and recursive self-improvement is not inevitable,” the authors write.
In his June essay, “Policy on the AI Exponential,” Amodei calls for stronger government oversight of the most advanced AI development. His proposal calls for independent testing, public reports about risks, and legal powers that would allow governments to stop or discourage releases that pose catastrophic dangers. Under Anthropic's October draft bill, a release could be blocked only if a model is judged to pose an immediate risk of killing or injuring more than 50 people or causing more than $1 billion in damage.
In September, Amodei outlined a three-stage plan in an essay titled “We Must Pace the Frontier.” The plan starts with independent evaluators working inside AI companies. It then moves to cooperation among democratic countries and, ultimately, to agreements involving both democratic and authoritarian governments. Anthropic argues that controlling the pace of advanced AI requires cooperation across the industry and around the world, along with safety measures that individual AI labs can take themselves.
Anthropic's Frontier Safety Roadmap, whose latest entries are dated October 19, sets December 15, 2026, as the target date for a policy proposal on how governments can support greater access for evaluators who work inside AI companies.
By January 15, 2027, Anthropic plans to publish a policy framework for shaping the pace of AI development or explain why it has decided not to publish one. The roadmap warns that automated AI research “could accelerate progress faster than governance or safety measures can keep up.”
The company acknowledges that its plans may change. It says it will publish updates explaining whether it meets its goals.