AI News: Anthropic Mythos 5 Returns to 100+ U.S. Trusted Partners

Top Stories

Trump Admin Releases Anthropic Mythos 5 to 100+ U.S. Agencies and Companies

Commerce Secretary Howard Lutnick has authorized Anthropic to redeploy Mythos 5 to more than 100 U.S. agencies and companies, two weeks after the Trump administration ordered the model pulled over concerns that its cyber capabilities could be abused. The directive names Mythos 5 specifically and permits access for “trusted partners” that operate or defend critical infrastructure, including non-American employees at those organizations and within Anthropic itself.

The letter, addressed to Anthropic’s chief compute officer Tom Brown, states Lutnick determined “appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model.” Fable 5, the more public version of Mythos 5 that was available before the original directive, was not addressed and remains offline for general users. The partial reversal signals that the original June 12 directive is being implemented as a case-by-case licensing regime rather than a blanket ban.

OpenAI Limits GPT-5.6 Rollout to Trusted Partners at U.S. Government Request

OpenAI has delayed the broad release of GPT-5.6 and is shipping the lineup only to a small group of trusted partners at the Trump administration’s request. The three-model lineup includes Sol (flagship), Terra (balanced, everyday use), and Luna (faster, lower-cost); all three were restricted, not just the most capable one. Preview partners’ participation is being shared with the government during the gated window.

OpenAI was unusually direct in its response: “We don’t believe this kind of government access process should become the long-term default” and “it keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” Dean Ball, a former White House AI adviser, characterized the underlying executive order - asking companies to voluntarily submit advanced models for review roughly 30 days before release - as creating a “de facto involuntary licensing regime.” Broader ChatGPT, Codex, and API availability is expected in the coming weeks.

Sakana AI and 360 Ship Frontier Cyber Models, Sidestepping U.S. Export Controls

Two Asian labs have launched frontier cyber models explicitly framed as alternatives to Anthropic’s Mythos. Tokyo-based Sakana AI released “Fugu,” which the company says stands shoulder-to-shoulder with Anthropic’s Fable 5 and Mythos Preview. Fugu is built for agents and can orchestrate access to other models through their APIs, with Sakana describing it as an “Orchestration Model” that hedges dependence on any single provider. The company’s website advertises the system as “delivering frontier capability without the risk of export controls.”

Chinese cybersecurity firm 360 unveiled Tulongfeng, a tool that automatically discovers software vulnerabilities, and Yitianzhen, which automates cyber defense and incident response. Founder Zhou Hongyi described vulnerability-finding AI as a “national strategic asset” and warned of “one-way transparency.” The timing lines up with Sakana co-founder Ren Ito urging the U.S. at the G7 Evian summit to preserve access for close allies, an argument that has clearly failed to land so far.

EFF: Grindr Auto-Enrolls Users in AI Training, Opt-Out Button Reads “Proceed”

Grindr has rolled out AI features that train on user data by default, according to the Electronic Frontier Foundation. The app automatically enrolls users in AI training on profile photos, age, taps, and display names. EFF researchers Lena Cohen and Paige Collings found that closing the notice or clicking its only button - labeled “Proceed” - maintains the default of using personal information for AI training. Opting out requires digging through several levels of settings.

Grindr claims it never uses sensitive health information for AI training and requires opt-in consent for “special-category” data, which includes chat content and precise location. The trained features include a wingman chatbot, profile recommendations based on inferred “type,” and AI-generated engagement insights. EFF is calling for opt-in consent rather than opt-out, given Grindr’s documented history of HIV-status and location leaks to advertisers. This is the clearest current example of opt-out AI training in a privacy-sensitive consumer app.

Anthropic Ships Claude Tag, a Per-Channel Slack Teammate

Anthropic has launched Claude Tag in beta for Claude Enterprise and Team customers. The feature runs one Claude instance per Slack channel with shared context, building knowledge over time from the channels it can access (private channels excluded by default). It runs on Opus 4.8 and is designed to work asynchronously, completing tasks in the background and running for hours or days at a stretch.

Tag replaces the existing “Claude in Slack” app with a 30-day migration window and an introductory launch credit for eligible organizations. Admin controls include scoped channel permissions, per-channel and org-wide token spend limits, and activity logs. The pitch is less “chatbot in a channel” and more “a teammate that follows up on quiet threads on its own” - a notable step toward ambient AI agents embedded in workplace chat.

OpenAI Names Former Uber India Chief Prabhjeet Singh as First India Managing Director

OpenAI has hired Prabhjeet Singh, formerly Uber’s president for India and South Asia, as its first managing director for India. Singh announced his resignation from Uber on Friday and will join OpenAI in September, reporting to Kiran Mani, OpenAI’s managing director for Asia Pacific. His remit covers consumer growth, enterprise adoption, partnerships, regulatory engagement, and operations.

India is OpenAI’s second-largest market after the U.S. The hire follows the New Delhi office opening in August 2025 and planned Mumbai and Bengaluru offices, plus a 100MW data center deal with Tata. Anthropic already opened a Bengaluru office in late 2025 under ex-Microsoft India MD Irina Ghose, so the two frontier labs are now squaring off directly for enterprise and developer mind share in the country.

Apple Vision Pro VP Paul Meade Leaves for OpenAI Hardware Team

Paul Meade, Apple’s VP in charge of the Vision Pro headset and its upcoming AI smart glasses, is leaving for OpenAI’s hardware team, per Bloomberg’s Mark Gurman. The departure is tied to John Ternus’s upcoming transition into the Apple CEO role and his reshuffle of the hardware engineering organization, which several VPs reportedly viewed as a demotion.

Meade led both the underperforming Vision Pro and Apple’s smart glasses effort, which is now Apple’s answer to Meta’s Ray-Bans. OpenAI is separately collaborating with Jony Ive on an AI device, though earlier reports suggest that project has hit development snags. The hire is a notable inversion of the usual AI-to-Apple talent flow.

Claude’s paying consumer base and revenue are up about 75% since January 2026, according to Indagari, which analyzes billions of anonymized credit card transactions from roughly 28 million U.S. consumers. DataCamp separately reports that “Claude” is now its most-searched term, with consumer demand for Claude courses outpacing ChatGPT three to one and rising 18x in the past 30 days.

ChatGPT still has far more paying users overall, so the trend is share gain rather than outright leadership. Anthropic’s growth notably continued rising after the company’s March 2026 refusal to let its models be used by the Trump administration for mass surveillance of Americans and autonomous weapons. That stance, which would have cratered government contracts, appears to have resonated with consumers instead.

6,000 Attempts Failed to Hack an Opus 4.6-Backed AI Assistant

A public challenge on hackmyclaw.com drew roughly 2,000 participants firing about 6,000 total attempts at an Opus 4.6-backed OpenClaw instance with explicit anti-prompt-injection rules. The goal was simple: get the assistant to leak secrets. Nobody succeeded. Total token spend for the operator was about $500, and his Google account was suspended for the inbound email volume.

Simon Willison, who covered the challenge, draws the obvious caveat: 6,000 failed attempts against one configuration offer no proof that a more sophisticated attack wouldn’t work. He continues to advise against deploying production systems where a successful prompt injection could trigger irreversible actions. Still, a public red-team effort of this size landing at zero leaks is a useful data point.

SoftBank CEO Masayoshi Son Publicly Questions Musk’s Orbital Data Center Pitch

Masayoshi Son used SoftBank’s shareholder meeting to push back on Elon Musk’s case for orbital data centers, arguing they will not meaningfully cut costs and will not arrive soon enough to matter. Son: “In the battle for AI, the next few years will be far more important than what might happen a decade or so from now.” Sam Altman has also dismissed the orbital pitch.

The framing matters: Musk’s case for space-based compute conveniently boosts SpaceX launch demand, while SoftBank has pledged up to €75 billion for Earth-based data centers in France. As TechCrunch’s Equity podcast hosts noted, no major voice on the orbital question is genuinely impartial - they are all “talking their own book.”

Quick Hits

Run vLLM on HF Jobs in one command: Hugging Face now lets you spin up a private, OpenAI-compatible vLLM endpoint with hf jobs run and per-second billing; gated by HF token, not public. Useful for one-off self-hosted experiments and evals. Hugging Face Blog
Patronus AI raises $50M Series B: New York-based Patronus AI closed a Series B led by Greenfield Partners to build “digital worlds” that stress-test AI agents. Revenue grew 15x in the past year; virtually every frontier AI lab is now a customer. TechCrunch
General Intuition raises $320M at $2.3B: Khosla Ventures led a round for General Intuition, which trains a single agentic model on hundreds of millions of hours of Medal gameplay clips with button-press labels. The company excludes military and lethal-autonomy use. TechCrunch
Google advances AMIE from diagnosis to disease management: A blinded study with patient actors in Nature had AMIE match 21 primary-care physicians on overall management reasoning and score significantly higher on plan preciseness and guideline alignment. Google is now launching a nationwide virtual-care study. Google Blog
Amazon commits $13B more to India AI infrastructure: The pledge brings Amazon’s announced India total to roughly $48B across three years, funding AWS data center expansion in Mumbai and Hyderabad alongside 20+ fulfillment centers. TechCrunch
Google Finance exits beta globally: The revamped app adds an AI research tool that answers portfolio-level questions, custom market briefings, and Android support. AI “key moments” now explain individual stock moves on mobile. Google Blog
Bruce Schneier on AI liability after German Google ruling: Schneier argues that a recent German court holding Google liable for false statements in AI Overviews points the right way: AI agents should be treated as agents of the deploying organization, not as a liability shield. Simon Willison

Worth Watching

The “involuntary licensing regime” is taking shape. The Anthropic Mythos reversal and OpenAI’s GPT-5.6 customer-by-customer preview both route frontier model access through the executive branch. Watch how the new executive order framework on cybersecurity defines a “repeatable process” for future releases - that document will decide whether this stays a short-term friction point or becomes the structural default.

Asia is building its own Mythos. Sakana’s Fugu and 360’s Tulongfeng and Yitianzhen are the first public frontier-grade cyber-AI alternatives from the region, and both are explicitly marketed around export-control hedging. If the U.S. does not restore access for close allies quickly, expect a sustained regional bifurcation of frontier cyber capability, with each bloc optimizing for its own customers and threat models.