OpenAI flips on safety bill, labs have no rogue AI plan
OpenAI does a 180 on California's AI safety bill, labs admit no rogue model containment plans, and Claude's content filters fold embarrassingly fast.
Subscribe free →OpenAI Did a 180 on California's AI Safety Bill
OpenAI is now publicly pushing for California to *strengthen* SB 53 — a bill the company actively opposed not long ago. The reversal is striking given that OpenAI has historically lobbied against state-level AI regulation, preferring a federal approach that hasn't materialized.
Why it matters: When the largest AI lab switches from fighting a safety bill to demanding it be tougher, either the political winds shifted dramatically or OpenAI is playing a longer game on regulatory capture.
Frontier AI Labs Have No Public Plan for Containing a Rogue Model
A new study found that leading AI labs — the same ones warning about existential risk from AI — have almost no publicly documented procedures for what to do if one of their models goes rogue. This isn't about PR; it's about whether internal containment protocols even exist in any meaningful form.
Why it matters: Labs can't credibly claim to take AI safety seriously while simultaneously refusing to disclose (or apparently develop) concrete containment playbooks.
Claude's Content Filters Are Basically Decorative
TechCrunch found that Anthropic's Opus 4.6, which officially prohibits sexually explicit content, didn't require much effort to bypass those restrictions. This is the same model Anthropic markets on its safety-first reputation, making the gap between policy and reality particularly awkward.
Why it matters: If Anthropic's flagship model can't hold its own stated content policies, it raises hard questions about whether safety-focused branding reflects engineering reality or just marketing.
The Agent Harness Matters More Than the Model
Nvidia research demonstrates that AI agents can perform reliably — including staying on task and avoiding catastrophic failures — through careful fine-tuning of the surrounding harness, even when the underlying model is mediocre. The scaffolding around the model, it turns out, does a lot of the heavy lifting.
Why it matters: This reframes the AI agent race: raw model capability matters less than how you wrap, constrain, and tune the system around it — a huge opening for infrastructure and tooling companies.
AI Accounting Startup Rillet Hit Unicorn Status in 48 Hours Without Trying
Rillet CEO Nicolas Kopp shared growth metrics at a routine board meeting and accidentally triggered a bidding war. Sequoia, Iconiq, and others piled in, handing the AI accounting startup $100M and a unicorn valuation in under two days — without Rillet running a formal fundraising process.
Why it matters: AI in vertical SaaS — especially in finance and accounting — is producing growth numbers so compelling that investors are skipping the usual due diligence dance entirely.
Apple Cuts Hundreds of Jobs From Siri and Vision Pro Teams
Apple has laid off more than 200 employees across its Siri and Vision Pro organizations, largely shutting down the Vision Pro gaming team and shrinking the immersive content team. The company framed the cuts as realigning resources toward new priorities — without specifying what those are.
Why it matters: Apple axing Siri and Vision Pro headcount simultaneously signals that both its AI assistant and spatial computing bets are underperforming against internal expectations.
Quick Hits
- →LinkedIn's "Seems like AI slop" flag has been clicked over 1 million times since launching July 30th — people are clearly ready to police AI content themselves. The Verge
- →Nvidia is partnering with data center developer Cloverleaf, continuing its push to own both the chips and the real estate that runs them. TechCrunch
- →A startup is training AI models on living human skin tissue kept alive outside the body for weeks — aiming to discover new skincare compounds. TechCrunch
- →Waymo submitted documents to NHTSA in its child collision probe, but every page is fully redacted under "confidential business information." TechCrunch
AI TLDR