Anthropic Asked for a Transparent Process. It Got a Friday Evening Phone Call.

The AI company had publicly called for government oversight that is fair, clear, and grounded in technical facts. What arrived was none of those things.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://open.substack.com/pub/johanosteyn/p/anthropic-asked-for-a-transparent?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

At 5:21 in the afternoon on Friday 12 June 2026, Anthropic received a letter from US Commerce Secretary Howard Lutnick. The letter, written with the involvement of the Bureau of Industry and Security, cited national security authorities and instructed Anthropic to immediately suspend all access to its latest models — Claude Fable 5 and Claude Mythos 5 — for any foreign national, whether inside or outside the United States, including Anthropic’s own foreign national employees. No written technical disclosure accompanied the directive. No evidence was provided of a harmful outcome. No process existed through which Anthropic could challenge the instruction before complying. Within hours, both models were offline for every customer, in every country, across the world.

What makes this moment significant is not the shutdown itself. It is the gap between the shutdown and the process Anthropic had explicitly and publicly asked for — a statutory mechanism for government oversight that is, in the company’s own words, transparent, fair, clear, and grounded in technical facts. What arrived on Friday evening was none of those things.

CONTEXT AND BACKGROUND

To understand what happened on 12 June, it is necessary to understand what Anthropic had done in the months before it. In April 2026, Anthropic launched Project Glasswing — a structured, government-collaborative framework for managing its most capable AI models responsibly. Rather than releasing its most powerful system immediately to the general public, Anthropic made Claude Mythos Preview available only to a small group of cyber defenders and critical infrastructure providers, in direct partnership with the US government. The explicit purpose was to develop safeguards robust enough to eventually make Mythos-level capabilities safe for broader release.

By June 2026, Anthropic concluded those safeguards were ready. The models had been red-teamed for thousands of hours by internal teams, the UK AI Safety Institute, and multiple external private organisations. A public bug bounty produced no universal jailbreaks in over 1,000 hours of testing. The solution was a two-tier release: Claude Fable 5 for the general public, with conservative safety classifiers blocking the most dangerous capability categories, and Claude Mythos 5 for Glasswing partners with some classifiers lifted. More than 95 per cent of Fable 5 sessions would involve no classifier fallback at all. The architecture was designed to make safety and broad access compatible rather than competing.

Three days after launch, the Commerce Department issued its directive. The stated basis was a jailbreak — a method of bypassing Fable 5’s safeguards. Anthropic reviewed what it believes to be the basis of the directive and concluded the vulnerability is a narrow, non-universal jailbreak that exploits only previously known minor vulnerabilities, provides no capability uplift beyond what is already available from other publicly deployed models including OpenAI’s GPT-5.5, and is the kind of non-universal jailbreak that Anthropic had explicitly stated was likely present in any frontier model and had factored into its defence-in-depth strategy from the start. Anthropic stated it had not received disclosure of a concerning non-universal jailbreak that led to a harmful result.

INSIGHT AND ANALYSIS

The governance failure the Fable 5 directive reveals is not Anthropic’s. It is the absence of the thing Anthropic had been asking for. In its public policy statements, Anthropic has consistently argued that government should have the ability to block unsafe AI deployments, as part of a statutory process that is transparent, fair, clear, and grounded in technical facts. The Friday evening directive was none of those things. It arrived verbally, without written technical evidence, without a defined standard against which the jailbreak was being assessed, without a process for Anthropic to respond before compliance was required, and without any consideration of whether the same standard was being applied to competitors whose models carry the same vulnerability.

I have previously written about the deeper question of what happens when AI intersects with governance and lawmaking, and the accountability and legitimacy questions that follow when those processes lack the transparency that democratic institutions depend on. The Fable 5 directive is a precise illustration of that problem in reverse: not AI participating in governance without accountability, but a government acting on AI without any of the procedural safeguards that make governance legitimate.

The political backdrop makes the directive harder to read as a purely technical safety intervention. In February 2026, the Trump administration barred Anthropic’s products from federal agencies after Anthropic sought guardrails on Pentagon use of its technology. Trump posted publicly that Anthropic had made a disastrous mistake trying to strong-arm the Department of War. Anthropic sued. A federal judge ruled in Anthropic’s favour, but the case continues in Washington. Despite that dispute, the Financial Times reported last week that the NSA was using Mythos to conduct offensive cyberattacks. On 2 June, Trump signed an executive order directing federal agencies to gain early access to frontier AI models on a voluntary basis. Anthropic launched Fable 5 and Mythos 5 seven days later. The directive arrived three days after that.

Read in sequence, the pattern suggests the directive was less a considered safety intervention than a political instrument — deployed in an ongoing dispute between a company that has maintained an independent safety position and an administration that has consistently treated that independence as an obstacle.

IMPLICATIONS

The consequences of the shutdown extended well beyond Anthropic’s US customer base. Because Anthropic has no technical mechanism to verify the nationality of individual users, complying with the instruction to block all foreign nationals required disabling both models entirely for every customer in every country. Businesses, researchers, developers, and institutions across Africa, Europe, Asia, and the rest of the world lost access to tools they had integrated into their operations — not because of anything they had done, not because of any finding about how they were using the models, and not through any process that involved them or their governments in any way.

This is the AI sovereignty argument made concrete and specific. Every organisation outside the United States that operates on US-hosted AI infrastructure is subject to having that infrastructure removed, without notice, without recourse, and without appeal, by a domestic US political decision they had no part in. That is not a hypothetical risk to be managed in a long-term strategic plan. It is an operational reality that materialised on a Friday evening in June 2026 and resolved only when Anthropic and the US government reached an accommodation that had nothing to do with the interests of users in Johannesburg, Nairobi, Lagos, or anywhere else.

For South African boards and executives, the Fable 5 shutdown joins a growing body of evidence about what dependency on US AI infrastructure actually means in practice. South Africa’s Draft National AI Policy, gazetted in April 2026 - then withdrawn - introduces a risk-tiered governance framework that classifies consequential AI systems as high-risk and requires traceable lines of accountability. That framework cannot protect South African organisations from the operational consequences of a US export control directive. The accountability it requires cannot be traced to a letter from a Commerce Secretary that arrived at 5:21 on a Friday afternoon.

CLOSING TAKEAWAY

Anthropic built the most rigorous pre-launch safety framework in the AI industry. It partnered with the government, which then overrode it. It called for a statutory oversight process that the government that issued the directive has not created. It complied with an instruction it publicly disagreed with, describing it as a misunderstanding, because it had no choice.

The lesson for boards and executives is not that Anthropic was wrong to ask for a transparent process. It was right. The lesson is that the transparent process does not yet exist — and that until it does, the governance of the world’s most powerful AI systems will continue to be shaped by the kind of opaque, politically inflected, Friday evening decisions that the companies building those systems, and the users depending on them, have no reliable way to anticipate or resist.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net