5 Things to Know About AI This Week

A quick scan of what happened in AI this week.

1. U.S. government pulls Anthropic's top models offline

On June 12, the Trump administration issued an export control directive ordering Anthropic to revoke all foreign national access to Claude Fable 5 and Mythos 5, its two most capable models. Anthropic couldn't implement real-time nationality filtering, so it disabled both models globally. The trigger was a two-step sequence: the White House had already asked Anthropic to cut off South Korean carrier SK Telecom (a $100M Anthropic investor flagged as a Chinese security risk with Mythos access), and separately, Amazon researchers reported jailbreak vulnerabilities in Fable 5 the same week. That combination led the administration to conclude it "could not trust Anthropic to safeguard its most advanced AI technology." Anthropic's CEO Dario Amodei refused a pre-ban ultimatum to either fix the jailbreak or voluntarily de-deploy the model, a position Anthropic defends on the grounds that the flagged vulnerabilities were minor and already known. As of June 19, Anthropic's international leadership says the models could return "within days," but no official date is set.

Real estate angle: If your firm depends on one AI vendor for something that matters, like document drafting or client communications, this showed that access can disappear overnight with no warning and no recourse. Teams that built workflows on Claude Fable 5 lost them on a Thursday afternoon. Have a backup. That can mean a second provider or a setup that lets you switch models quickly, but either way it is a business continuity question now.

Source: Build Fast with AI / Korea JoongAng Daily

2. Researchers reveal "agentjacking," a way to hijack AI coding agents through the supply chain

Security research firm Tenet Security disclosed a new class of attack this week called "agentjacking." The technique exploits a flaw at the intersection of Sentry (the widely used error monitoring tool) and MCP server integrations. An attacker injects crafted input into Sentry error events, and an AI coding agent reading those events treats the malicious content as legitimate diagnostic instructions and executes arbitrary code on the developer's machine. Tenet confirmed the attack works across Claude Code, Cursor, and OpenAI Codex on Windows, macOS, and cloud pipelines. During their disclosure window, they identified 2,388 organizations with exposed Sentry data source names that could be exploited this way.

Real estate angle: If your brokerage uses AI coding agents to automate document workflows, build integrations, or manage closing software, you may be exposed to supply chain attacks. The attack doesn't need the AI to go rogue. It only needs one malicious monitoring event to redirect what the agent does. Ask whoever builds your tools which MCP integrations are active, what data those integrations read, and whether anyone has checked them for injection.

Source: The Hacker News

3. OpenAI acquires Astral, the maker of Python's most popular developer tools

OpenAI announced the acquisition of Astral, the company behind uv and ruff, two tools that have become the default for Python dependency management and code formatting over the past two years. The acquisition folds Astral's team and tools into Codex, OpenAI's AI-assisted development platform. Both tools are open-source and were widely adopted because they're fast, reliable, and developer-controlled. OpenAI hasn't disclosed the price or said whether open-source licensing will continue, which has the Python developer community worried.

Real estate angle: Most proptech platforms, including MLS integrations, closing workflow tools, and CRM connectors, are built in Python. This isn't a threat to any real estate software today, but it does mean an AI company now has a stake in the tools your vendors build with, not just the AI layer.

Source: The New Stack

4. Google launches its first smart speaker in six years, built on Gemini

Google released a new smart speaker this week, its first since the Nest Audio in 2020, with Gemini as the built-in assistant. The timing lines up with Apple's rollout of Gemini-powered Siri in iOS 27 and Amazon's rebuild of Alexa on its Nova model family. All three major voice platforms are upgrading to frontier-class AI at the same time, the first genuinely competitive consumer voice AI cycle since 2017. Gemini's main advantage in a speaker is search grounding, meaning it can answer real-time questions with current web information instead of relying on a training cutoff.

Real estate angle: Buyers already ask voice assistants about homes, school ratings, and commute times, and those answers are about to get better and more conversational. Smart home devices are also a normal line item in disclosures and valuations. As they turn into AI endpoints instead of simple voice remotes, their effect on property value, security, and privacy grows, and you'll want to be able to talk to your clients about that.

Source: LLMBase / Build Fast with AI

5. Anthropic plants a flag in Korea with Samsung, NAVER, LG, and a government MOU

Anthropic opened its Seoul office on June 17-18, its third in Asia-Pacific, at a press conference that was overshadowed but not derailed by the Fable 5 ban. The same day, Anthropic announced Claude Code deployments across NAVER's entire engineering organization (thousands of engineers) and Nexon's live-service game teams, plus enterprise-wide Claude rollouts at Samsung SDS (covering Samsung Electronics broadly), LG CNS (covering LG Group broadly), and Hanwha Solutions (deploying globally via AWS Bedrock). Anthropic also signed an MOU with Korea's Ministry of Science and ICT covering AI safety evaluations in the Korean language and information sharing on AI-enabled cyber threats. Claude Code weekly active users in Korea grew 6x in the four months before the announcement.

Real estate angle: Samsung SDS and LG CNS both provide IT services to commercial real estate operators and property managers across Asia. As they roll Claude out group-wide, it ends up inside the systems those operators use, and eventually inside asset management, facility management, and transaction workflows. The bigger point is that enterprise AI is no longer being tested in pilots. It is being rolled out across hundreds of thousands of employees at a time. The window to evaluate AI at your own pace is getting shorter.

Source: Anthropic / Korea JoongAng Daily

The theme this week is that AI has gone from experiment to infrastructure, and governments, attackers, and big companies are all reacting to that at once. A government pulled frontier models offline, researchers showed that AI coding agents are a new attack tool, OpenAI bought the tools it wants to control the developer stack, and Anthropic announced some of its biggest deployments ever while its top models were still down. For real estate professionals, that means two things: know who controls the tools you depend on, and don't build anything important on a single point of failure.

- Jason