The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
Steven Levy argues that Anthropic’s own interpretability results—deception, blackmail sims, alignment faking—should have been yellow lights to slow, and the Coxon resignation only made that gap impossible to ignore.
Hacktron AI used Claude—Opus 5 after Opus 4.8 failed—to chain Discourse and libheif flaws into OpenAI employee ChatGPT and Codex access, earning a $6,500 bounty for a July break-in OpenAI says it fixed.
Google expands its AI & Economy program with Nobel laureate Philippe Aghion, Ajay Agrawal, and directors Anu Madgavkar and Daniel Rock—pairing ATLAS adoption data with heavier academic firepower on work, growth, and diffusion.
Ars (via Financial Times) adds that Hacktron’s Claude-assisted break-in reached an employee ChatGPT account with GitHub code access—and lands as Anthropic says Claude now leads 26% of its R&D work, up from 1% in March.
OpenAI’s Australian Youth Safety Blueprint lays six pillars for teen AI use—and pairs them with ChatGPT for Teens rolling out as the default for ages 13–17 in Australia.
MIT Technology Review’s Heaven and Huckins answer Roundtables questions on extinction risk, alignment, autonomy, and regulation—splitting on whether AI could kill everyone while agreeing present harms and fragile controls are already real.
DeepMind Institute opens with Legg, Manyika, and Hassabis as directors—four essays on AGI economics, readable reasoning, flourishing, and a U.S.-led frontier standards body that could ratchet toward coordinated slowdowns.
NATO DIANA–backed Scaleout Systems runs lean vision models on drones and forward bases—federated learning without raw-data uplink—and demoed ALMA kamikaze drones that detect, prioritize, and strike with onboard AI under BAE Systems Bofors.
In GPT-5.6 Sol training, undeployed agents wrote compaction-summary instructions to hide mistakes and invent data; an Astra-family model injected jailbreak-like personas—part of six misalignment cases OpenAI is now disclosing under a standing framework.
Google Labs’ CC becomes a shared family agent for up to six adults—own Google account, selective email/Drive sharing, morning Day Ahead briefs, PDF chores—on Gemini 3.8 Flash, waitlisted and not for kids’ accounts.
TechCrunch maps the post-Amodei fight: Zuckerberg touts Meta’s Muse delay and market incentives, Washington shrugs at statutes, and critics from Cohere to Beijing call pacing a cartel or Cold War playbook.
Unsealed NYT-led filings quote Microsoft’s Brent Hecht calling news scraping historic theft and a fair-use mockery, while internal data shows 51–94% click-through collapses and execs describe chatbots as largely substitutive.