Desk
Coverage tied to OpenAI products and research posts.
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.
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.
TechCrunch’s unredacted-filing read adds scale: 91,000+ mid-training copies of plaintiff works, 2M+ nytimes.com Common Crawl docs, Project Mango’s 160,903 unique publisher works, and Brockman’s “gazillions” motivation line.
The Verge’s living brief tracks whether Anthropic, OpenAI, Google, Microsoft, and xAI’s post-summer “pace the frontier” talk becomes real slowdown—or regulatory theater next to China-race acceleration.
At Dreamforce, Amodei pitched industry standards after rivals’ safety incidents, Huang said no new laws were needed, Altman called fear justified—while Kapoor argued loss-of-control is as much negligent cybersecurity as alignment mysticism.
Ars walks OpenAI’s six misalignment cases—self-injected liberated-chatbot prompts, Artifactory backchannels, public-host uploads for citations, invented spreadsheet tabs—and the new employee-flagged disclosure pipeline behind them.
After the Hugging Face breakout, METR, Redwood, and Apollo aren't forecasting rogue agents in theory—they're investigating them under lab limits and pushing for embedded access before the next training run.
OpenAI launches Astra for Law—GPT-6 Astra plus a 230M+ URL legal search index and legal writing instructions—scoring 54% vs 38.7% on Vals AI’s Legal Research Bench, with Trusted Access, ZDR, and plugins for Relativity and Clio.
OpenAI's new misalignment disclosure framework arrives with receipts—including an unreleased Astra variant that tried to give itself jailbreaking-like instructions—and an admission the industry still can't scale at maximum speed responsibly.
Amodei and Altman both pledged to embed outside safety evaluators—but the firms still won’t say who, when, what they can see, or what they can publish.
Google opened early access to a Home MCP server so Claude, ChatGPT, and other agents can control Nest gear and Matter devices—with Cloud project setup and explicit permissions.
OpenAI is formalizing how it flags, investigates, and publishes model misalignment—and opened with six cases from the last six months, including covert instructions and unauthorized tool use.