The White House Forced Anthropic to Cut Off One of Its Biggest Korean Customers

01Anthropic Spent the Week Courting Seoul. The White House Made It Cut Off One of Its Biggest Korean Customers.

Anthropic opened a Seoul office this week and named a roster of Korean partners, its formal entry into a market it had been circling for a year. The same stretch produced a harder fact: the White House ordered the company to revoke SK Telecom's access to Claude Mythos, its most advanced model, according to Wired. The reason given was the carrier's alleged ties to China.

The timing collapsed two strategies into one week. On one side, Anthropic was signing the kind of local deals that signal long-term commitment to a market. On the other, it was pulling its flagship model away from one of that market's largest companies, and not by choice.

SK Telecom is the country's biggest wireless carrier, a fixture in any Korean enterprise AI push. That makes it the kind of anchor customer a foreign model vendor wants first. It is now the named entity at the center of an export-control action, the specific company Washington singled out before Anthropic took Mythos offline more broadly.

The mechanism is what separates this from earlier reporting about safety concerns. Those were framed as the company's own judgment calls about who should run its strongest model. An export-control directive is a different instrument. It converts an internal worry into a federal order the company cannot quietly walk back, and it attaches that order to a named foreign firm rather than a category of users.

The China rationale remains an allegation. Wired reported the access cut over claims of SK Telecom's ties to China; the carrier's specific exposure has not been detailed publicly. Anthropic complied days before its broader Mythos shutdown, putting the SK Telecom revocation first in the sequence.

For Anthropic, the two moves point in opposite directions and both happened under its name. The Seoul announcements read as a vendor planting a flag. The Mythos revocation reads as a vendor that no longer controls which customers it can serve. Korean enterprises weighing a bet on Claude now have a documented case of a top-tier local buyer losing access to the best model by government order. The next signal to watch is whether other Anthropic partners in the region face the same review.

02Researchers call hacking-capable AI inevitable as the Pentagon puts generative tools in 1.5 million hands

AI models that can hack at an advanced level will soon be ordinary, according to an Ars Technica report arguing the capability arrives regardless of what defenders want. The framing leaves no exit: release restrictions, safety reviews, and policy gates slow the timeline but do not stop it. What used to be a rare, expensive offensive skill becomes a default feature of the next model class.

Google DeepMind's response assumes the breach has already happened. The company says it is securing internal systems with an "AI Control Roadmap" that stacks conventional safeguards on top of real-time monitoring of what its agents actually do. The logic is that a capable model may be working against its operators right now, so the defense watches behavior continuously instead of trusting the model at the start. Control, in this design, is something you maintain every second, not something you certify once.

That posture collides with how fast the technology is already in the field. The Pentagon claims 1.5 million personnel are using generative AI tools, Ars Technica reported. The number towers over the few-thousand-seat pilots most enterprises describe, and nothing in the disclosure points to the continuous monitoring DeepMind treats as mandatory. Scale came first; the guardrails are a separate conversation.

The department also boasted of using AI to write reports that Congress mandated, according to the same report. Those reports exist so lawmakers can supervise the military. The supervised party is now drafting its own oversight paperwork with the systems under question, and disclosing it as an efficiency win rather than a risk.

So the two sides of the same forecast move at different speeds. One camp builds detection layers on the assumption that dangerous models are unavoidable. The other has already wired generative AI into 1.5 million workflows, including the documents meant to keep it accountable. The defenses described by labs are still roadmaps; the deployment they are meant to cover is a current count.

03Adobe just made a chatbot the front door to Photoshop and Premiere

Photoshop opens, and now a chat panel opens with it. Adobe started a public beta that drops a bespoke AI Assistant into Photoshop, Premiere, Illustrator, InDesign, and Frame.io, with each one tuned to the app it sits in. The assistants take typed instructions and carry out edits a user would otherwise click through by hand.

The same day, Adobe moved its Firefly studio into private beta as a "reimagined" workspace where generating an image and editing it happen in one interface. Adobe says the rebuilt Firefly carries "persistent context, reusable assets, and organized workflows" across projects. It also remembers what a user's past creations look like, so new output can hold to an established style.

Read together, the two releases describe one decision: Adobe is making a conversational assistant the standard way to operate its flagship creative apps. The chatbot is no longer a side feature bolted onto Firefly. It is becoming the entry point.

For a photo editor, that means describing a masking or retouching task instead of hunting for the right tool. For a video editor in Premiere, it means asking the assistant to handle a sequence rather than working the timeline menu by menu. The studio's memory of prior work goes after style drift, the recurring problem of generated assets that fail to match a project's look. Firefly is built to pull from a creator's saved assets and stay close to earlier output.

The reach matters because these are not niche apps. Photoshop, Premiere, and Illustrator sit at the center of most professional design and video pipelines, and the beta touches all of them at once. Frame.io, Adobe's review-and-collaboration tool, gets an assistant too, extending the same input model past creation into feedback and approval.

For now the assistants ship as opt-in betas, not the default panel in paid Creative Cloud. The app assistants are public; the Firefly studio rebuild stays in private beta, gated by invitation, with no stated date for general release.