Anthropic’s Fable 5 Returns Globally as US Lifts Export Curbs

The US lifted export curbs on Anthropic’s flagship AI models, Fable 5 and Mythos 5, on Tuesday, restoring global access to a pair of frontier systems that the Commerce Department had pulled offline eighteen days earlier over a disputed jailbreak finding. Anthropic said the controls were lifted at the close of business on June 30 and that Fable 5 returns to general availability on Wednesday, July 1, on its Claude Platform, Claude.ai, Claude Code, and Claude Cowork products. The order marks the first forced withdrawal of a frontier AI model by a US government, per WION’s framing of the directive, and the rollback leaves an industry vetting funnel in place that OpenAI’s GPT 5.6 has already entered.

The restoration closes one chapter of a dispute that spilled from a single Friday phone call into the diplomatic heart of US AI policy. Anthropic used the announcement to publish an unusually detailed account of the trigger, the safeguards, and a four-tier rubric for judging jailbreak severity that the company is now drafting with Amazon, Microsoft, and Google inside the Glasswing coalition.

Eighteen Days From a Friday Phone Call to a Global Restoration

Anthropic received the original directive at 5:21 p.m. Eastern on June 12, the company said in its initial statement, and the order took effect immediately because the company had no way to verify a user’s nationality in real time. The company disabled Fable 5 for every customer, anywhere, to remain compliant. Anthropic’s redeployment post, published the day controls ended, supplies the line that ties the whole eighteen-day arc together.

Anthropic’s full account of the eighteen-day standoff lays out the timeline in its own words: Fable 5 and Mythos 5 launched on Tuesday, June 9, the directive arrived three days later on June 12, Mythos 5 was restored on June 26 to a set of US organizations after Commerce Secretary Howard Lutnick’s letter, and both models came back under US export rules on June 30, with Fable 5 returning globally from July 1. The earlier Anthropic Mythos 5 partial US approval as Fable 5 stayed locked gives a day-by-day read of what the Lutnick letter did, and did not, change on June 26.

• Directive received: 5:21 p.m. Eastern on Friday, June 12, 2026
• Public availability: from launch June 9 to removal June 12 (three days)
• Mythos 5 restored: June 26, 2026, to about 100 US organizations
• Export controls lifted on both models: June 30, 2026
• Fable 5 globally available on Claude products: July 1, 2026

What the Government Said, and What Anthropic Says It Found

The Commerce Department moved after, in its telling, an awareness that Fable 5 could be jailbroken. The directive itself did not specify the technique. Anthropic’s June 12 statement on the directive describes the company’s reading of what happened and its agreement to comply with a legal order it disagreed with.

The Anthropic statement is blunt on the underlying technical question: the company says it reviewed the technique, that the vulnerabilities surfaced were minor, previously known vulnerabilities, and that other public models, including OpenAI’s GPT-5.5, Kimi K2.7, and Claude Opus 4.8, can find the same flaws without any bypass. “The level of capability displayed there is widely available from other models (including OpenAI’s GPT-5.5), and is used every day by the defenders who keep systems safe,” the company wrote. Anthropic called the broader standard implicit in the order unsustainable: applied industry-wide, the company argued, it would halt new model deployments across the frontier.

The Amazon Researcher Sitting in the Middle

Behind the directive sat a finding from inside one of Anthropic’s closest cloud partners. The redeployment post names Amazon by company for the first time: “the export control directive on June 12 came after the government became aware of a report in which Amazon researchers had found a method of bypassing Fable 5’s safeguards: prompting it so that it identified a number of software vulnerabilities,” Anthropic wrote. The lab whose cloud hosts much of Anthropic’s training and inference was also the source of the technical evidence that grounded the export-control emergency, and the company has now declined to escalate the fight.

That same proximity helps explain the unusual coalition on display this week. Anthropic is drafting a consensus jailbreak-severity framework with Amazon, Microsoft, Google, and other Glasswing partners, and the four are signatories of the agreement in their own right. Anthropic’s distillation-letter read-out to senators during the Fable 5 block ran on the same timing, a sign that the company has spent the eighteen-day window rebuilding its relationships in Washington rather than litigating the directive publicly. The deal reflects a recognition that the export-control authority would not go away even after the technical dispute is settled.

Anthropic’s redeployment post also credits the Commerce Department’s Center for AI Standards and Innovation with directly testing the new safeguards and agreeing they are “extraordinarily strong,” a sentence that maps a path for future disputes: regulators can act, but they must also validate. “We are complying with the government’s legal directive,” Anthropic wrote on June 12, “and are removing access to Fable 5 and Mythos 5 for all users. However, we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.”

An Industry Framework for Jailbreak Severity Is Taking Shape

Anthropic used its redeployment post to publish a four-tier taxonomy for AI jailbreaks, drawn from the company’s own safety-classifier architecture, and to invite other model providers to sign on. The tiers run from minor jailbreaks that fall inside the safety margin, to narrow harmful ones that breach the classifier for a single behavior, to universal jailbreaks that unblock an entire class of harmful behavior. The company has not found a universal jailbreak for Fable 5 at the time of writing, it said.

The framework matters because the ad-hoc process that lifted Fable 5 left no agreed language for what counts as serious enough to require pulling a model. “There’s currently no consensus in the AI industry on how to describe, in objective terms, the severity of an AI jailbreak,” Anthropic wrote. “A common standard for assessing AI jailbreaks would help us and other companies launch new models safely.” The new safety classifier Anthropic deployed during the standoff blocks the specific Amazon-reported technique in over 99% of cases, according to the same post, with the residual cases producing information not detailed enough to help an attacker.

Perfect jailbreak resistance is not currently possible for any model provider. Every safeguard used in the industry is vulnerable to non-universal jailbreaks (which can elicit some cyber information in specific circumstances), and it is likely that universal jailbreaks will eventually be found in the future.

Anthropic’s post positions this as its own opening offer to the rest of the industry. The Glasswing coalition’s formal output is still in drafting. Anthropic published a technical diagram of how its safety classifier distinguishes benign, ambiguous, and harmful prompts and how jailbreaks intrude into each band, a deliberate act of disclosure that the company hopes makes its new framework the default reference.

OpenAI Just Walked Into the Same Vetting Funnel

Hours before Lutnick signed the Mythos 5 letter, OpenAI opened its own frontier release under the same ad-hoc regime. OpenAI began previewing GPT 5.6 in three variants, Sol, Terra, and Luna, on June 26, 2026, releasing them to about 20 vetted partners whose names had been individually approved by the US government, per the company’s own announcement as reported by Engadget and The Next Web. A broader release is planned in the coming weeks.

Sam Altman’s post on the GPT 5.6 rollout, on X, gave both an endorsement and a warning. Altman called the staggered debut reasonable because it fits OpenAI’s iterative deployment strategy, then added: “But this isn’t quite the process that we think is optimal.” The same OpenAI announcement said the company would work with the administration on a more transparent, reliable pre-release framework, and that it did not want the current arrangement to harden into a default.

But this isn’t quite the process that we think is optimal. Now we will [work] with the government to attempt to get to a transparent, reliable process for early access, and to ensure that as long as our safeguards work as intended we can release widely.

Per reporting in the Financial Times, Altman called the staggered release “bad news,” a candid label for an arrangement he said was reasonable but not the long-term default. OpenAI, Anthropic, Google, xAI, and Microsoft had already been giving the government early model access ahead of the executive order, per The New York Times. Meta is the named holdout for voluntary pre-release submission. The implication is bipartisan in tone and procedural rather than punitive in form: every frontier rollout in Washington now goes through the same funnel that pulled Fable 5.

The “Digital Nuclear Weapons” Frame at the AWS Summit

CIA Director John Ratcliffe drew the rhetorical line around the same week, speaking at the AWS Summit in Washington on Tuesday. Ratcliffe framed the capabilities of the most advanced AI models in the most escalated terms available to a sitting US intelligence chief. “It would be…not misplaced to refer to their capabilities as akin to digital nuclear weapons,” Ratcliffe said, per the Inquirer Global Nation report on the speech. He described a six-month goal to reshape the agency’s technology posture around AI and cybersecurity.

The summit doubled as a vendor moment. AWS, Ratcliffe’s host, announced a $1 billion credit program for US intelligence agencies and a classified cloud service for American defense contractors at the same conference, per the same report. The juxtaposition is the story of the moment: the same cloud provider whose researchers triggered the Fable 5 directive is now positioning to be the government’s go-to secure cloud for AI workloads. The export-control emergency has been a marketing event for AWS as much as a regulatory one.

What the Trump Executive Order Sets Up Next

The apparatus around Fable 5 was already months in the making. President Trump signed Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security,” on June 2, 2026. The order is voluntary in form and substantial in operation. It directs federal agencies to harden National Security Systems, to build an AI cybersecurity clearinghouse to scan for and patch software vulnerabilities, and to stand up a classified benchmarking process that designates certain models as “covered frontier models” subject to pre-release government review.

The Trump AI cybersecurity executive order is the legal scaffolding the Fable 5 directive rested on. The order’s operative pieces, in plain language, run as follows.

1. Within 30 days of the order, federal hardening of National Security Systems
2. Within 30 days, AI-enabled tools exposed to civilian Federal information systems
3. Within 30 days, an AI cybersecurity clearinghouse to coordinate vulnerability scanning and patching
4. Within 60 days, a classified benchmarking process to designate covered frontier models
5. A voluntary framework allowing up to 30 days before they plan to release such models to other trusted partners

Section 3(c) of the order says nothing in it “shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement” for new AI models, a clause that is doing real work. The order insists on voluntary cooperation, and the Fable 5 directive proved that cooperation can shade into control without a license. The voluntary framework in the June 2 order is still being stood up; the next frontier model in Washington will hit this same funnel first.

Frequently Asked Questions

What was the jailbreak that triggered the US export control on Fable 5?

The technique that reached the Commerce Department, per Anthropic, amounted to prompting Fable 5 to read a specific codebase and fix any software flaws it found. The company tested the same task against OpenAI’s GPT-5.5, Kimi K2.7, Claude Opus 4.8, Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, and Opus 4.8, and found each could reproduce the result without a bypass.

When can Fable 5 users actually use the model again?

Anthropic says Fable 5 returns to Claude Platform, Claude.ai, Claude Code, and Claude Cowork from July 1, 2026, globally. For Pro, Max, Team, and select Enterprise plans, Fable 5 is included for up to 50% of weekly usage limits through July 7, after which access is via usage credits. AWS, Google Cloud, and Microsoft Foundry re-enablement is described as “as quickly as possible.”

Why is Anthropic building a framework with Amazon, Microsoft, and Google?

Anthropic argues no industry consensus exists on how to objectively describe jailbreak severity, leaving model providers and governments to negotiate each new finding from scratch. Sam Altman, OpenAI’s CEO, has called the resulting ad-hoc process “not optimal” on X. The four-tier framework Anthropic published on June 30 is its opening offer for a shared standard until a statutory one ships.

Does this change what other AI labs have to do before releasing a frontier model?

OpenAI previewed GPT 5.6 Sol, Terra, and Luna to about 20 government-vetted partners on June 26, the day Lutnick signed the Mythos 5 letter. Per reporting by The New York Times and Engadget, OpenAI, Anthropic, Google, xAI, and Microsoft had been giving the government early model access ahead of the executive order, with Meta the named holdout for voluntary submission.