What just happened? The shutdown of Anthropic's latest public AI model has ended, following a dispute that prompted Washington and the tech industry to take a closer look at AI security. The Trump administration and Anthropic have reached an agreement to restore access to Fable 5, the company's latest general-purpose model, after security concerns prompted its shutdown. The pause followed Amazon researchers' discovery of a way around the model's safeguards.
The workaround exposed gaps in the model's guardrails. Fable, a public-facing version of Anthropic's more capable Mythos system, is designed with restrictions to prevent harmful uses, including assistance with cyberattacks. Those guardrails are critical for releasing the model publicly.
Anthropic responded by adding a new layer of safeguards. According to the company, the technique flagged by Amazon now fails about 99% of the time. In the remaining cases, Anthropic said the outputs either rely on publicly available information or do not provide meaningful help to a bad actor. Earlier, the company had also rerouted certain risky prompts to a less advanced model as a stopgap.
The dispute underscored a familiar problem in AI: the more capable the systems get, the harder they are to control. Anthropic acknowledged as much in its discussions with other major tech firms, saying it is "probably impossible" to make any model jailbreak-proof. The company is now working with Amazon, Microsoft, and Google to establish a shared way of evaluating these kinds of vulnerabilities and deciding how to respond.
Anthropic executives, including Chief Compute Officer Tom Brown, were in direct talks with federal agencies to argue that the model was safe to restore. The Center for AI Standards and Innovation, a government testing unit, evaluated both the original safeguards and the updated protections as part of that process.
The standoff marked a sharper turn in Washington's approach to AI oversight. Until recently, oversight had been relatively hands-off. That has changed as national security and cybersecurity concerns move closer to the center of AI policy. A new executive order asks companies to give the federal government access to their models 30 days before public release.
The administration had imposed restrictions on foreign access to Anthropic's models, including access for foreign-born employees working in the US, prompting the company to shut the model down entirely. Those limits are now expected to be lifted following the agreement.
The dispute also highlighted Amazon's dual role in Anthropic's business. The company has already invested $13 billion and could put in $20 billion more. At the same time, it was Amazon researchers who brought the vulnerability to light.
Tensions between Anthropic and the federal government have surfaced before, including disagreements over military use of AI and export restrictions on advanced chips. President Trump, who earlier described the company as a "radical-left, woke company," struck a different tone after the negotiations, saying of CEO Dario Amodei, "He responded very responsibly."
More broadly, the episode has left some uncertainty about how AI models will be approved and monitored. Industry executives and policymakers say the recent actions have disrupted the usual model approval process, but they expect things to settle once the new executive order is fully in place.
Anthropic said it will begin restoring access to Fable 5 on Wednesday. "We're grateful to our users for their patience, and to everyone who worked with us on redeploying," the company said.