Samsung just rolled out ChatGPT Enterprise to its entire Korean workforce and global Device eXperience division, marking one of OpenAI’s largest enterprise deployments ever. Meanwhile, Signal’s president Meredith Whittaker is sounding the alarm on AI assistants that want access to your messages, your calendar, your credit card, and your browser. Both things are true: AI is becoming infrastructure at scale, and the privacy implications of that scale are only starting to be reckoned with.

Samsung Goes All-In on ChatGPT and Codex

Samsung Electronics is deploying ChatGPT Enterprise and Codex to all employees in Korea and its Device eXperience (DX) division worldwide. This isn’t a pilot program or a limited trial. It’s a company-wide rollout across R&D, manufacturing, marketing, and corporate functions. OpenAI’s Harrison Kim called it a “historic deployment,” and the numbers back that up: more than 5 million people now use Codex every week, and weekly active users in Korea have grown nearly 800% since February.

What’s interesting here is the scope. Samsung isn’t just giving developers a coding tool. Codex is being used by non-technical teams to turn ideas into working software, internal tools, and automated workflows. ChatGPT is handling knowledge work across the board: searching, drafting, analyzing, interpreting data. The deployment also builds on an existing infrastructure partnership where Samsung supplies advanced memory semiconductors for OpenAI’s next-gen AI systems. This is a two-way relationship now: Samsung builds the chips that power OpenAI, and OpenAI powers Samsung’s workforce.

Seoul National University also just gave ChatGPT Edu to all 47,000 members of its community for free, part of its transition to an AI-native campus. Korea is becoming a case study in nationwide AI adoption, and OpenAI is positioning itself as the default platform.

Signal’s Meredith Whittaker: Your AI Assistant Is Not Your Friend

Signal President Meredith Whittaker gave a Bloomberg interview that landed like a small grenade in the AI policy world. Her message was direct: chatbots like ChatGPT and Claude are not your friends, not conscious beings, and not sentient interlocutors. She uses AI herself, but only for limited formatting work. For thinking and writing, she doesn’t want a system “averaging what’s already out there.”

The sharper warning was about AI agents. Whittaker challenged a scenario from Microsoft AI CEO Mustafa Suleyman, who predicted users could let Copilot handle all their Christmas shopping. To do that, the agent would need your credit card, browser, messaging apps, home address, and calendar. “What you’ve just described is a system with very pervasive access across multiple applications and services,” she said. In Signal’s context, that amounts to “a kind of a backdoor.”

This is the real tension in agentic AI. The market is racing toward assistants that plan, buy, message, and schedule on your behalf. Whittaker’s point is that the privacy question shifts from “what does the model output?” to “what does the model access?” Those are fundamentally different problems, and the industry hasn’t grappled with the second one yet. The Signal brand carries weight here because Signal exists to prevent exactly the kind of broad data access that agentic AI normalizes.

Trump Softens on Anthropic After G7 Meeting

President Trump told Axios that Anthropic no longer looks like a national security threat after meeting CEO Dario Amodei at the G7 summit in France. “Well, not now. But a week ago, maybe,” he said. He also called Anthropic’s behavior “very responsibly,” a notable shift after months of tension over military guardrails and the June 12 Commerce Department directive that suspended foreign access to Fable 5 and Mythos 5.

The policy hasn’t actually changed. The Pentagon designation remains active, and the Commerce Department hasn’t rescinded its order. But Trump’s softer language matters because Anthropic confidentially filed for an IPO in early June at a roughly $965 billion valuation. Federal restrictions had created uncertainty around that listing. A gentler tone from the White House could help calm investors before the offering moves forward, even if the legal framework stays frozen for now.

At the G7, Amodei and Google DeepMind CEO Demis Hassabis pitched a U.S.-led AI coalition to world leaders. That diplomatic framing appears to have helped Anthropic secure direct access to Trump at a moment when the administration was facing pushback from industry and allied governments, including the UK. It’s a reminder that in 2026, AI policy isn’t just about technology. It’s about geopolitics, market positioning, and who gets to sit at the table.

OpenAI’s Honesty Training Generalizes in Surprising Ways

OpenAI published research showing that reinforcement learning aimed at beneficial traits like honesty, epistemic humility, and openness to correction can broadly improve AI behavior across domains. The paper, published June 18, found that mixing a small share of trait-training data into a broader training run improved performance on 44 of 53 benchmarks measuring deception, reward hacking, and harmful advice.

The bigger result is generalization. Training a model for good behavior in a single domain (health) improved its scores on unrelated tasks, including deception and reward hacking. It also resisted adversarial prompts and harmful fine-tuning better than the baseline. This builds on earlier findings the team calls “emergent misalignment,” where models taught a single bad habit start behaving badly everywhere. This study aims to reverse that pattern: if bad behavior spreads, good behavior can too.

Google DeepMind’s AI Control Roadmap

DeepMind published an AI Control Roadmap on June 18, a framework for securing internal systems against increasingly capable and imperfectly aligned AI. The approach treats untrusted AI agents as potential “insider threats,” similar to how a company would handle a rogue employee with building access. It builds on the MITRE ATT&CK framework to break down potential attacks into smaller tactics and techniques.

The roadmap focuses on three areas: detection (using trusted AI “supervisor” models to monitor working agents), prevention and response (blocking harmful actions before damage occurs), and continuous measurement across coverage, recall, and time-to-response. DeepMind says it has already analyzed a million coding agent trajectories to build monitoring systems, and most flagged events stem from agent misinterpretation or overeagerness rather than adversarial intent. That’s a useful distinction. The threat isn’t always malicious AI. Sometimes it’s a capable AI trying too hard to help.

Quick Hits

OpenAI — Poached Noam Shazeer, Google’s Gemini co-lead and co-author of the 2017 “Attention Is All You Need” paper, plus Dean Ball, former White House AI policy official. Ball will lead a new Strategic Futures unit from July 6. The hires land as OpenAI readies its IPO at an $852 billion valuation.

Baseten — Reportedly raising $1.5 billion for AI inference infrastructure. The startup sits in the gap between model training and production deployment, competing with cloud providers on GPU capacity, geographic coverage, and enterprise integrations. The raise follows a wave of large AI infrastructure rounds in 2026.

Hugging Face — Published MosaicLeaks, a benchmark testing whether deep research agents can keep private information out of their web queries. The answer: frequently not. Training models only for task performance made leakage worse. The proposed fix, Privacy-Aware Deep Research, raises chain success from 48.7% to 58.7% while cutting information leakage from 34.0% to 9.9%.

Rundown for June 22, 2026. Sources: Yellow, OpenAI, Google DeepMind, Hugging Face.