The White House Wants Anthropic to Build Jailbreak-Proof AI. Security Experts Say That’s Impossible.

The White House Wants Anthropic to Build Jailbreak-Proof AI. Security Experts Say That’s Impossible.

In my first piece on this story, I covered the what: the US government killed Anthropic’s best models with export controls. In the second piece, I covered the who: Amazon CEO Andy Jassy personally triggered the crackdown. Today, we know the why — and the next demand from the White House may be technically impossible to meet.

SK Telecom: The Company at the Center

The “South Korean telecommunications company” the White House suspected of China ties is SK Telecom. Wired confirmed the identification this week (Wired, June 17). SK Telecom is South Korea’s largest wireless carrier and invested $100 million in Anthropic in 2023.

Here’s where it gets complicated. SK Telecom is part of the SK Group conglomerate, which has extensive business in China — semiconductors, energy, the works. The company had a historical joint venture with China Unicom, a state-owned carrier, dating back to 2004 with a $1 billion investment in 2006. The first Trump administration restricted US investment in China Unicom in 2021, and the FCC proposed barring interconnection in April 2026.

SK Telecom’s own China revenue was only $1.9 million in 2024 — seven employees. But it’s the conglomerate ties that raised red flags. The White House asked Anthropic to revoke SK Telecom’s access. Anthropic complied immediately.

Then, on June 17, Anthropic opened a Seoul office and still lists SK Telecom as a partner. That’s a bold move.

The Impossible Demand

This is the core of the escalation. White House officials told Wired that Fable 5 can only come back online when Anthropic “fixes” the jailbreak. The jailbreak involves asking Fable to “fix this code” — getting it to find security weaknesses. The NSA concluded there ARE ways to disable the guardrails.

But here’s the problem: security experts say this is technically impossible. As one expert put it: “Your AI is either highly skilled at writing secure code or it is not. You cannot draw a distinction between offensive and defensive capability at this level.” The same vulnerabilities are “easily identified by Opus 4.8 and GPT-5.5.” You can’t patch a capability without removing the capability.

Anthropic’s position: jailbreak effects are “minimal,” no universal jailbreak exists, and perfect jailbreak resistance is “not currently possible for any model provider.” The government is demanding something that may not be solvable.

May Be Illegal

Politico reported that legal experts are questioning the legal authority (Politico, June 18) behind these export controls. Kevin Wolf, a former senior Commerce official with 25+ years of experience, said flatly: “I don’t know what the legal authority is for that.”

Matthew Borman, former deputy assistant secretary of Commerce, noted this is the first time the government has tried to restrict access to a thing unrelated to access to the underlying technology. Commerce hasn’t published the rule in the Federal Register. If left unchallenged, the precedent could allow the government to impose export controls on ChatGPT, Gemini, and every frontier model.

The Cascade Effect

The export controls are already rippling through Wall Street. JPMorgan Chase has stopped its Hong Kong staff from accessing Anthropic’s AI models, according to the Financial Times. The restriction was triggered by the wording of Anthropic’s usage terms in its licensing agreement. Goldman Sachs did the same thing in April, removing Claude from its approved tools for Hong Kong bankers.

Meanwhile, Trump said Wednesday that negotiations with Anthropic are “going fine.” But the power dynamics behind those negotiations are shifting fast.

The Power Shift

Axios reported that David Sacks is stepping back from AI policy and Sriram Krishnan is departing. Howard Lutnick, the Commerce Secretary, is now leading — and his letter is what triggered the export controls in the first place. Scott Bessent at Treasury was the point of contact when Amazon raised concerns. Susie Wiles, the Chief of Staff, was receptive to Bessent’s concerns about Mythos’s impact on the financial sector.

There’s friction too: National Cyber Director Sean Cairncross thinks Treasury overstepped. Bessent’s allies think Cairncross hasn’t met the moment. Cairncross’ head of policy, Thomas Lind, plans to leave — further depleting the technical expertise driving these decisions.

What Happens Next

Anthropic dispatched senior staff to DC for a meeting Monday with Commerce officials and the National Cyber Director’s office. It ended without resolution. The company is complying publicly while pushing back privately — opening a Seoul office, still listing SK Telecom, calling the action a “misunderstanding” that “does not adhere to” principles of transparency and fairness.

The legal challenge is building. The technical demand may be unsatisfiable. The political coalition behind the controls is fracturing. And Anthropic is quietly daring the government to escalate further.

I’ll update this post as the story develops. If you’re just catching up, start with the original breakdown and the Amazon angle.

Submit a Take

Your email address will not be published. Required fields are marked *