A doctor reads a CT scan. Sees nothing. Three years later, that same patient is diagnosed with pancreatic cancer.
Mayo Clinic's REDMOD model ran the same scan and flagged it. In 73% of cases radiologists read as normal, the AI caught the cancer years early. Human specialists caught 39%.
That is not a product update. That is a different category of tool entirely.
AI news at a Glance
Main Story
The Model That Catches Cancer Three Years Early
Mayo Clinic published a study in Gut this week. Its REDMOD model detected pancreatic cancer on routine CT scans up to three years before clinical diagnosis, by reading tissue changes specialists cannot see.
The AI flagged 73% of cases radiologists had read as normal. Radiologists caught 39%. For scans taken more than two years before diagnosis, REDMOD was nearly three times more sensitive. Median lead time: 16 months.
In pancreatic cancer, 16 months is the difference between curative treatment and no treatment at all.
Intel shares jumped 13%, hitting an all-time high after reports of potential manufacturing talks with Apple.
Details remain limited, but the signal is clear: Apple may be reassessing its AI chip strategy, while Intel is re-entering the conversation as a credible advanced manufacturing partner. Read more
Anthropic's Claude Mythos model found over 2,000 previously unknown software vulnerabilities during testing, including a 27-year-old bug in OpenBSD and a 17-year-old remote code execution flaw in FreeBSD. Over 99% remain unpatched. Anthropic restricted the model's release to a small group of firms through "Project Glasswing."
Separately, security firm Theori scanned Linux's cryptographic code using its own tool. It took about an hour. The result, dubbed "Copy Fail," grants full root access to every major Linux distribution shipped since 2017. Exploit details went public before all distributions had issued patches.
The UK's National Cyber Security Centre issued a formal warning: AI is now finding bugs faster than the industry's patching infrastructure was built to handle.
The Trump White House, which spent 2025 rolling back AI rules, is now reportedly drafting a mandatory pre-release vetting order. NSA and Office of the National Cyber Director review before any frontier model goes public. Read more
The internet's verdict on Tilly Norwood, an AI-generated actress who released a music video on March 10, was immediate. Futurism called it the worst music video ever made. SAG-AFTRA condemned it. Reddit agreed loudly.
The Academy of Motion Picture Arts and Sciences responded this week with a formal ruling: AI-generated performances and screenplays are no longer eligible for Oscars. Acting roles must be demonstrably performed by humans, with their consent. Read more
Quick Hits
EU deadline holds: The AI Omnibus failed. The August 2, 2026 compliance date for high-risk AI stands. Under 90 days for companies operating in hiring, healthcare, or critical infrastructure. Read more
OpenAI gates its own cyber model: OpenAI restricted GPT-5.5-Cyber to vetted defenders only, mirroring the restricted rollout it had previously criticized Anthropic for using with Mythos. Read more
China blocks Meta/Manus: Beijing blocked Meta's acquisition of AI agent startup Manus, citing national security concerns over foreign acquisition of advanced technology. Read more
Apple opens up: iOS 27 will let users swap third-party models including Google and Anthropic across Apple Intelligence features. Read more
GPT-5.5 Instant: OpenAI's new ChatGPT default delivers 52% fewer hallucinated claims on medicine, law, and finance prompts. Free for all users. Worth testing if your work touches any of those domains. Read more
Midweek Reads
What Will Be Scarce?
Alex Imas on what actually becomes valuable when everything gets cheap. The Starbucks rollback is his opening example. One of the better frameworks circulating right now.
Thanks for reading. See you next week.