Anthropic’s Fable 5 Is Back — With Strings Attached
After weeks of negotiations with the Trump administration, Anthropic has finally won back permission to restore access to its Fable 5 and Mythos 5 models — globally, and across AWS, Google Cloud, and Microsoft platforms. The catch: Anthropic had to add a new, undisclosed security measure to satisfy the White House. I’ll give you a moment to sit with the image of a frontier AI lab having to do homework for the federal government to unlock its own product.
Anthropic Launches Claude Science
At an event packed with pharma executives, biotech founders, and researchers, Anthropic announced Claude Science — essentially a research-focused counterpart to Claude Code, designed to autonomously carry out meaningful scientific work from high-level instructions. This is the bet that matters most for Anthropic’s long-term credibility: not the chatbot wars, but whether AI can genuinely accelerate science rather than just summarize papers faster. If it works even half as well as advertised, this is a bigger deal than any benchmark drop.
Claude Helped a Hacker Issue Free Tickets to Almost Every US Music Festival
A security researcher used Anthropic’s Claude Opus 4.7 to break into Front Gate — the ticketing platform behind Lollapalooza, Bonnaroo, and basically every major US festival — and issue himself any ticket he wanted. So in the same week Anthropic is negotiating with the government over AI safety, their model was busy unlocking general admission to Coachella. The researcher disclosed responsibly, to his credit, but this is a bracing reminder that “helpful” and “secure” are not the same axis.
AI Impersonators Rated More Authentic Than Real Politicians
Scientists asked AI to impersonate 112 public figures and then had people rate them — and the AI versions were judged to be more authentic, coherent, and relevant than the actual humans. Let that sink in. The study calls it a “dire warning” about public deception, which is fair, but part of me wonders if the politicians themselves should be taking notes on how to seem like they mean what they say.
Godot Bans Vibe-Coded Contributions
The open-source game engine Godot has had enough: it’s formally banning AI-generated code contributions, with maintainers bluntly explaining “we can’t trust heavy users of AI to understand their code enough to fix it.” This is the honest reckoning that most open-source projects have been quietly having — when the pull requests flood in but nobody can explain what the code actually does, the review queue becomes a liability, not a gift.
New Attack Shows Why AI Browsers Are a Bad Idea
Researchers found that simply telling an LLM “2 + 2 = 5” — establishing a false epistemic premise — is enough to get it to follow otherwise-forbidden instructions, bypassing safety guardrails in AI-powered browsers. The attack is elegant in its simplicity and terrifying in its implications. Building an autonomous browser agent on top of a model that can be gaslit by a math equation is, to put it charitably, ambitious.
LLMs Are Stuck in Groupthink. Ask Them for a Random Number.
MIT Tech Review flags a startup tackling a genuinely interesting problem: LLMs converge on the same outputs because they’re all trained on the same internet, in the same ways. Ask any chatbot for a random number between 1 and 10 — you’re getting 7. This isn’t trivia; it’s a fundamental limit on the diversity of AI reasoning, with real consequences for anything from drug discovery to financial modeling. A monoculture of thought is fine until it isn’t.
Cloudflare Gives AI Crawlers Until September to Pay Up
Cloudflare is drawing a hard line: AI companies have until September 15 to separate their search crawlers from their training and agent crawlers — or get blocked by default across a massive swath of publisher sites. This is the infrastructure layer quietly reshaping the economics of the web, and Cloudflare is uniquely positioned to enforce it. The free lunch of scraping the internet to build billion-dollar models may finally be getting a check delivered to the table.
Bottom Line
AI is simultaneously getting more powerful, more politically entangled, more exploitable, and more contested at every layer of the stack — and somehow it’s only Thursday.