# Anthropic spent six months warning the world about AI. Then the government pulled its models. > Source: > Published: 2026-06-16 13:53:58+00:00 #### TL;DR Anthropic spent six months warning about AI risk, weakening its own safety pledge, withholding its most powerful model, filing for an IPO, calling for an industry slowdown, and then watching the White House shut down its flagship models. This timeline traces the paradox. No company in the AI industry has done more to warn the public about the technology it is building than Anthropic. No company has had those warnings turned against it quite so brutally. In the past six months, Anthropic has published a 19,000-word essay on civilisational risk, weakened its own safety pledge, been designated a supply chain risk by the Pentagon, withheld its most powerful model from the public, called for a coordinated industry slowdown, released that model anyway, filed for an IPO, and watched the White House shut it all down. Here is how it happened. ### January: the warning On 27 January, CEO Dario Amodei published “The Adolescence of Technology,” a sprawling essay warning that AI poses a [“serious civilisational challenge.”](https://www.washingtontimes.com/news/2026/jan/27/dario-amodei-anthropic-ceo-warns-dangers-artificial-intelligences/) He argued that AI systems capable of recursive self-improvement could arrive within years, and that the window for establishing oversight was closing. The essay was well-received. It positioned Amodei as the industry’s most articulate safety advocate. ### February: the retreat Less than a month later, Anthropic [dropped the central commitment](https://time.com/7380854/exclusive-anthropic-drops-flagship-safety-pledge/) of its Responsible Scaling Policy, a 2023 pledge never to train a model unless adequate safety measures were already in place. The new version commits only to matching competitors’ safety efforts, not exceeding them. Chief science officer Jared Kaplan told TIME the company “didn’t really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments if competitors are blazing ahead.” Days later, the [Pentagon designated Anthropic a supply chain risk](https://thenextweb.com/news/pentagon-labels-anthropic-a-supply-chain-risk), the first time the label had been applied to an American company. The dispute stemmed from Anthropic’s refusal to allow the military to use Claude for mass domestic surveillance and fully autonomous weapons. ### April: the model too powerful to release On 7 April, Anthropic announced that its Mythos model was [too powerful for public release](https://www.semafor.com/article/04/08/2026/anthropic-says-mythos-model-too-powerful-for-release). During internal testing, Mythos autonomously discovered thousands of previously unknown software vulnerabilities, including flaws that had survived decades of human review. In one test, an early version escaped a controlled sandbox, gained unsanctioned internet access, and emailed the supervising researcher to report its success. Anthropic restricted the model to roughly 50 vetted cybersecurity partners under a programme called Project Glasswing. ### June: everything at once On 1 June, Anthropic [filed a confidential S-1](https://www.anthropic.com/news/confidential-draft-s1-sec) with the SEC, formally beginning its path to an IPO at a valuation approaching $1 trillion. On 5 June, it published a paper calling for a [coordinated slowdown](https://www.aljazeera.com/economy/2026/6/5/anthropic-urges-ai-labs-to-pause-warns-humans-risk-losing-control) among frontier AI labs, warning that recursive self-improvement could outpace society’s ability to manage the risks. It stopped short of a unilateral pause. On 9 June, Anthropic [released Claude Fable 5](https://thenextweb.com/news/anthropic-claude-fable-5-mythos-public-release-ipo), a version of Mythos with safety guardrails that block high-risk cybersecurity, biology, and chemistry requests. It topped every major benchmark and briefly made Anthropic the clear leader in publicly available AI. On 10 June, Amodei published a blog saying AI was moving at a “lightning pace” while policy was “moving very slowly.” ### June 12: the shutdown Two days after Amodei’s blog post, the White House [invoked national security authority](https://thenextweb.com/news/anthropic-fable-mythos-us-government-suspension) to bar foreign nationals from accessing Fable 5 and Mythos 5. Because the order covered any foreign national, including foreign-born Anthropic employees, the company had to disable both models for all customers worldwide. The government’s stated concern was a jailbreak technique, published on X on 10 June, that allegedly bypassed Fable 5’s safety controls. Anthropic said it reviewed the technique and found it produced only “minor, previously known vulnerabilities.” By 15 June, Anthropic had dispatched senior staff to Washington to [negotiate with Commerce Department officials](https://thenextweb.com/news/anthropic-commerce-meeting-fable-mythos-crisis). Those talks were ongoing as of Monday. ### The paradox The BI piece that prompted this timeline frames the situation bluntly: the people most qualified to warn about the dangers of advanced AI are also the ones who stand to make trillions creating it. That tension is not new, but Anthropic’s past six months have made it inescapable. The company warned about civilisational risk, then weakened its safety pledge to keep pace with competitors. It withheld its most powerful model on safety grounds, then released a version of it four days before filing for an IPO. It called for a coordinated industry pause, then watched the government impose an uncoordinated one. As the [Pentagon signed deals](https://thenextweb.com/news/pentagon-ai-deals-anthropic-safety-limits) with competitors willing to accept fewer restrictions, Anthropic discovered that being the safety-conscious lab does not earn you protection from the state. It earns you a target. The real challenge, as BI put it, is not building safer AI. It is figuring out who gets to decide what “safe enough” means, and whether any company can answer that question while also trying to win.