Visa lets ChatGPT spend your money as AI agents go autonomous

The week marked a major leap in AI autonomy. Visa wired ChatGPT to shop and pay on behalf of users, allowing AI agents to complete transactions at any Visa merchant without manual clicks. Anthropic released Claude Fable 5, its most powerful public Mythos-class model, achieving state-of-the-art benchmarks—especially in software engineering—while shipping with guardrails that block cybersecurity, bio, and chem queries (triggering in under 5% of sessions). Jeff Bezos came out of stealth with Prometheus, a $41B startup backed by JPMorgan, BlackRock, and Goldman Sachs, building an 'artificial general engineer' that ingests physical-world data to accelerate manufacturing of everything from jet engines to smartphones.

Security and regulation tightened in parallel. A self-replicating worm named Miasma compromised 73 repositories across four Microsoft GitHub organizations by injecting config files into open-source projects that auto-execute when developers open repositories in AI coding tools like Claude Code or Cursor. GitHub responded by planning to disable npm's auto-run install scripts by default in v12. Meanwhile, a German court ruled Google directly liable for its AI Overviews' content after false statements tied publishers to scams, stripping the traditional search-result shield. Anthropic broke with the White House, urging Congress not to preempt state AI laws without first passing rigorous federal safety tests for the most capable models.