A lawyer submitted fabricated witness testimony generated by ChatGPT in a murder appeal, prompting a judge to ask whether the attorney reads the news at all. This is the hallucination problem at its most consequential — a man's freedom potentially hinging on AI slop dressed up as legal record.
Cryptographer Matthew Green describes a two-part attack: a payload that hijacks an AI agent, which then carries that payload to the next agent — jumping across supposedly isolated sandboxes. This is the agentic security nightmare scenario, and it's no longer theoretical.
Anthropic's frontier red team finds Claude Mythos Preview successfully develops full control flow hijacks in 6% of trials — edging out GLM-5.3 at 4%. Capable models are becoming capable offensive cyber tools, full stop.
A new study identifies Cloudflare, Google, and Proton as dominant infrastructure providers for sites dedicated to non-consensual deepfake content. The pipes enabling the open web are the same pipes enabling mass sexual abuse imagery — and nobody is rushing to fix that.
Isenberg breaks down the DevDay launches he thinks will generate real money: Dots (OpenAI's personal agent platform), the Decisions API, and Sign In with ChatGPT. The last one is quietly the most strategic — OpenAI building identity infrastructure is a platform power grab, not a feature.
Dashboard cameras on postal trucks will scan roads, signs, and surroundings under the banner of "community safety." The most trusted institution in American neighborhoods is quietly becoming one of its most pervasive data collectors.
Mollick flags the coming collision nobody is prepared for: every company's customer service infrastructure is about to face a flood of AI agents — Dots, Muses, and their kin — negotiating on behalf of humans via voice and chat. The phone trees and friction-filled portals built to exhaust human patience mean nothing to a bot with infinite patience and zero embarrassment. This reshuffles consumer power in ways businesses haven't modeled yet.
Hassabis is retweeting data showing Gemini 4 Argon's hallucination rate at 15% — against GPT-6 Astra's 45% and Opus 5.5's 59%. If accurate, that's not an incremental improvement, it's a different category of reliability. Argon also just landed #1 on Text Arena. The three-way race Ethan Mollick flagged just got a new frontrunner.
Five private companies are now worth more than every tech IPO from 2015 to 2025 combined. a16z is making the case that the public markets have become almost irrelevant to where real value is being created — and that the IPO is just another financing event. This framing serves a16z's interests, but the underlying data is hard to argue with.
The through-line across today's feed is agents becoming load-bearing infrastructure for everything — commerce, law, security, and surveillance — faster than any institution is equipped to handle. The fake witnesses in the murder appeal aren't an anomaly; they're a preview of what happens when professionals outsource judgment to systems they don't understand. The lawyer didn't misuse ChatGPT — they used it exactly as marketed, and the guardrails weren't there.
Matthew Green's agent worm research and Anthropic's red team data land in the same week that OpenAI launches Dots — a persistent, always-on personal agent platform. The timing is not ironic, it's just the pace of 2026. Capabilities ship; security research follows at a sprint trying to catch up. Gemini 4 Argon's 15% hallucination rate is genuinely exciting, but a 15% error rate in an agent that autonomously negotiates, files, and acts is still a lot of errors at scale.
The USPS camera story and the deepfake infrastructure study belong in the same conversation. Surveillance and abuse don't require new technology — they require pointing existing, trusted, widely-deployed infrastructure in a new direction. Cloudflare didn't build a deepfake network; it built a CDN that deepfake networks use. USPS isn't a spy agency; it's a delivery service with continental reach. The architecture of the next panopticon is already installed. Someone just has to decide to use it.