Anthropic Ships Government Red-Team Model as Gemini Claims Chrome Tabs and Worldcoin Dangles Tinder Boosts

01Two months after Trump called it "Leftwing nut jobs," Anthropic hands Washington a red-team model

For nearly two months, the Trump administration has publicly attacked Anthropic. Officials called the company a "RADICAL LEFT, WOKE COMPANY," labeled its staff "Leftwing nut jobs," and described Anthropic as a menace to national security, according to The Verge. Federal buying channels reportedly closed off. Then Anthropic released Claude Mythos Preview, a cybersecurity model built for red-team work, and The Verge reports the administration's posture has started to shift.

Mythos is pitched as a tool for defensive security operators: finding vulnerabilities, simulating attacks, probing code. That is work federal agencies buy in volume. It is also the work an AI company trying to reopen Pentagon contracts would want to be visibly good at. The Verge reports the model has already softened at least some of the hostility coming out of the White House.

The warming is partial, and the underlying fight has not ended. Anthropic remains in a legal battle with the Pentagon over which AI uses it will accept, according to MIT Technology Review. The company has declined certain military applications even as the current conflict with Iran has pushed AI deeper into targeting and intelligence workflows, the report says. How the court rules will affect every frontier lab that has tried to hold a military-use line while courting federal revenue.

That refusal is the friction point. MIT Tech Review argues that the "human in the loop" framing both sides cite is functionally an illusion. The idea that a person always signs off on an AI-generated military decision breaks down once AI handles the intelligence analysis feeding that decision. Operators, according to the report, rarely have time or context to override what the system produces. That argument cuts against the safeguard Anthropic's military carve-outs are meant to preserve.

So Anthropic is offering Washington two things at once: a model strong enough for offensive-style security testing, and a set of red lines on battlefield use. Federal buyers, having spent two months branding the company a national security threat, now have to price its safety posture as an asset or a disqualifier. The Pentagon suit will set the terms, and the Iran conflict will keep raising the stakes while it runs.

02Chrome tabs and Google Photos are now Gemini's default context

Google rolled out two updates this week that wire personal data directly into its AI stack. AI Mode in Chrome can now read the content of the tab a user is viewing and open source pages in a side panel while Gemini answers. The Gemini app's new Nano Banana 2 image model pulls from Google Photos, generating pictures that include the user's family members, pets, and past vacation scenes.

Neither feature is pitched as a model upgrade. Both are pitched as context upgrades.

That framing marks a shift in what an AI assistant competes on. For two years the benchmark arms race dominated: parameter counts, MMLU scores, latency. Google's pitch here looks different. AI Mode's advantage comes from seeing the open tab. Nano Banana 2's comes from seeing years of Photos uploads. Neither capability requires a larger model.

The contrast with the current default is concrete. ChatGPT still starts each session cold unless a user manually opts into memory. A Gemini user signed into Chrome and Photos surrenders that cold start automatically, since both integrations are opt-in at the account level rather than per-query. Google says data stays within the user's account; the company did not publish retention windows for how long context is held for inference.

The switch is easy to flip on and harder to flip off. Personalized outputs degrade gracefully only while the model retains the context. Removing Google Photos access rolls back the feature set, not just the data pipe. Users evaluating the trade once, at setup, now determine what their assistant sees for every subsequent query.

The cost falls on people who already synced Chrome and Photos for convenience. Turning either off no longer costs them a cloud feature; it costs them the assistant that depends on it. Google's blog posts did not state whether personal context is used only at inference or retained for training.

03Stare into the orb, get five free Tinder boosts

A Tinder user in Japan or Argentina can now walk up to a chrome sphere called the Orb, let it scan their iris, and walk away with five free Boosts — the in-app perk that pushes their profile to the top of the stack. The trade is simple: prove you are a human being, get rewarded inside a dating app drowning in AI-generated profiles.

The Orb belongs to World, the identity project co-founded by OpenAI CEO Sam Altman. Tinder began testing the integration last year using World's facial-scanning hardware. It is now expanding to iris verification and rolling out across more countries, according to TechCrunch.

Tinder is the marquee name, but not the only one. World told reporters it is signing "multiple new partners" and positioning its verification credential, World ID, as infrastructure for any app that needs to distinguish humans from bots. Match Group, Tinder's parent company, has been piloting identity checks for years; the Orb gives it a cryptographic yes-or-no answer backed by biometrics the user cannot easily fake.

The pitch lands because the problem is real. AI-generated photos, chat replies, and entire synthetic personas have made dating platforms a laboratory for impersonation. Tinder needs something stronger than a selfie check. World is offering a credential that says "a verified human stood in front of an Orb" without revealing which human.

The irony sits underneath. The same technology wave that created mass-market generative AI — led in no small part by Altman's other company — is now producing demand for a separate layer of hardware whose entire purpose is to certify that the person on the other end is not a machine. World gets to sell both sides of the trade.

What began as a fringe crypto-adjacent experiment is quietly becoming plumbing. Every new partner adds a reason for the next app to adopt World ID, and every user who scans their iris adds a record to a biometric database run by a private company.