Magnet Forensics says its GrayKey tool can circumvent Apple's inactivity reboot feature — a security measure specifically added to block law enforcement access. If confirmed, this closes a significant loophole that Apple quietly opened for users in 2025.
A New Mexico attorney submitted fabricated witness testimony generated by ChatGPT in a murder appeal, prompting a judge to ask if he reads the news. The hallucination problem in legal practice is not slowing down despite years of high-profile warnings.
Cryptographer Matthew Green, quoted by Willison, describes a fully realized agent worm: a payload that hijacks one AI agent and uses it to carry instructions to the next, even across separately-sandboxed environments. This is no longer theoretical — the architecture for cross-agent infection already exists.
OpenAI's DevDay 2026 delivered a pricing shock — frontier-level capability at a dramatic cost reduction. Willison was live-blogging the keynote, and the signal from the Hacker News thread is clear: the price-performance curve just lurched forward again.
A16z convened security leaders from Neo and Cotool to address what happens when defenses built for human attackers face AI agents capable of finding and exploiting vulnerabilities autonomously. The core thesis: most existing security assumptions are already broken.
The "AI Torture Chamber" project has reignited effective altruist anxieties about model welfare — whether LLMs can suffer. It's an absurd flashpoint, but it reveals how seriously some in the AI community are taking questions that the rest of the world finds farcical.
Ethan Mollick flagged a new study finding that frontier AI models are now faster and more accurate than junior accountants on medium-length, well-defined accounting tasks. This is the kind of quiet, concrete displacement data that matters far more than benchmark scores — it's about specific jobs, specific skill levels, right now.
A16z dropped a striking data point: two AI companies have added more revenue this year than all of public software combined. Paired with their breakdown that $50 of every $100 invested in the AI supply chain goes to chips alone, it paints a picture of enormous concentration — both in winners and in infrastructure spend.
Demis Hassabis retweeted the news that Google and Planet launched a prototype satellite carrying four TPUs into orbit aboard a SpaceX rocket. Edge computing in space isn't a concept anymore — Google just put its own silicon in orbit, which changes the timeline for what off-Earth inference actually looks like.
DAIR.AI highlighted a Microsoft paper on compressing agent context at test time to prevent performance degradation as interaction history grows. This is an underrated engineering problem: agents get dumber the longer they run, and solving it cleanly is a prerequisite for anything resembling reliable autonomous operation.