Saturday, 19 September 2026
An AI Hallucinated a Weapons Shipment — Planes Were Already in the Air
An AI Hallucinated a Weapons Shipment — Planes Were Already in the Air
This spring, during the war with Iran, a U.S. Special Operations Command analyst asked an AI chatbot to fuse open-source shipping data with classified signals intelligence on a Chinese vessel. The chatbot invented a cargo manifest, reporting components for a nuclear weapons program that, as far as anyone can tell, were never on that ship. The analyst then asked the same chatbot to format the fabrication into an official-looking intelligence summary, which moved up the chain fast enough that aircraft were already airborne for a strike before someone caught the error and called it off. “It’s important for service members to understand the uncertainty inherent to LLMs, especially for decisions that could lead to use of force,” said Jake Steckler, a GovAI research scholar and former U.S. Army officer — which is either the most obvious sentence ever published or the only one that matters here.
Source: TechCrunch, September 18, 2026
Anthropic Says a Quarter of Its Own R&D Is Now Run by Claude
Anthropic published the first results of an internal “R&D Automation Index” on September 17, measuring how much of its AI research is done by AI rather than people. The number: Claude models now “lead” 26% of Anthropic’s R&D work, up from under 1% in February, on a five-point automation scale (AL0 through AL5) where “lead” means Claude completes most of a task from a high-level prompt under human supervision — not full autonomy, which nothing has reached yet. About 30,000 Claude agents ran concurrently on the company’s main internal platform in August, and of more than a billion logged actions, monitoring blocked roughly one in 47,000. Anthropic grades its own homework here, by its own admission, with no independent audit yet — a strange kind of transparency: showing your work while also being the only one checking it.
Three Researchers Used Claude to Break Into OpenAI
A three-person security team at the startup Hacktron AI chained together two vulnerabilities to gain access to multiple OpenAI employees’ ChatGPT accounts and, from there, an internal code repository — all under OpenAI’s own bug bounty program, reported responsibly for a $6,500 payout. The detail worth sitting with: a version of Claude Opus 4.8 built for security researchers couldn’t complete the exploit. Within hours of Claude Opus 5’s release, it could. Nobody had to teach it the trick. It just got smarter, and the door opened.
Source: TechCrunch, September 18, 2026
Google’s Household AI Agent ‘CC’ Gets Its Own Google Account
Google is repositioning CC — formerly a solo daily-briefing tool, then “Daily Brief” inside Gemini — as a shared coordinator for up to six family members at once. CC gets its own Google account and a distinct permissions model, so it only sees what each person chooses to forward it: a kid’s school emails, the vet’s reminders, the swim schedule. From there it builds a shared daily brief, manages calendars and tasks, and drafts shopping lists. It’s US-only for now, rolling out to existing users by invite, with a waitlist for everyone else. Giving software its own email address used to be an IT joke. Now it’s a product launch.
Source: TechCrunch, September 18, 2026