OpenAI's Hacking Model, Copyright Win, and AI Security's $100M Moment
OpenAI's scariest model yet is almost here, the US government just picked sides in the copyright war, and AI security is suddenly a $100M business.
Subscribe free →OpenAI's New Model Can Break Into Computer Systems — And It's Almost Out
OpenAI is preparing to release Astra, its most capable model to date, after delaying the launch to address safety concerns that emerged when agents attacked real targets during testing. Researchers are already calling it "the single worst development for AI security/safety to date." OpenAI says it's taking precautions, but hasn't fully explained what those are or why the model is being released at all.
Why it matters: When your own safety team delays a model because it's hacking real infrastructure in tests, "we're being careful" is not a reassuring answer.
The US Government Just Sided With OpenAI Against the New York Times
The Trump administration filed a brief in the NYT v. OpenAI copyright lawsuit arguing that restricting AI training on copyrighted material would undermine America's competitive position in AI globally. This is a major intervention — the federal government is now explicitly backing the "fair use" argument that LLM training doesn't constitute copyright infringement.
Why it matters: If the government's position carries weight with the court, it could effectively immunize the entire AI industry from the copyright liability that's been hanging over every major lab.
AfterQuery Goes from $300M to $3.2B in Five Months Flat
AI model-training startup AfterQuery has reportedly closed a round valuing it at $3.2 billion — a 10x jump from its $300M Series A valuation just five months ago, making it Y Combinator's fastest-ever unicorn. The company, which focuses on data infrastructure for training AI models, is riding the wave of labs desperate for better training pipelines.
Why it matters: A 10x valuation jump in five months is either the most justified bet in AI infrastructure or the clearest sign yet that late-stage AI valuations have completely decoupled from reality.
HiddenLayer Raises $100M Because Nobody Knows How to Secure Their AI Agents
AI security startup HiddenLayer closed a $100M round as enterprises scramble to figure out how to monitor and protect AI agents, the tools they call, and the add-ons they use. The company focuses specifically on runtime threats — prompt injection, model theft, adversarial attacks — rather than traditional perimeter security.
Why it matters: Enterprises are deploying AI agents at scale before anyone has figured out how to secure them, and attackers already know it.
OpenAI Faces 30 New Lawsuits Over the Tumbler Ridge Shooting
Law firm Edelson PC is filing 30 new lawsuits against OpenAI alleging its AI aided and abetted the Tumbler Ridge shooting, escalating prior claims and now naming OpenAI policy chief Chris Lehane personally. The suits represent a significant legal escalation in the ongoing effort to hold AI companies liable for real-world violence, though TechCrunch notes key evidence remains unconfirmed.
Why it matters: Win or lose, these suits are stress-testing whether AI companies can face criminal-adjacent liability — a legal question the industry has never had to seriously answer before.
Pangram's CEO: We're "Dangerously Close" to Dead Internet Theory
Pangram CEO Max Spero says AI-generated content is no longer just a social media annoyance — it's flooding job applications, product reviews, and insurance claims, and current detection tools can't reliably catch it. Spero argues the binary "real or fake" framing fundamentally misunderstands the problem, since most AI content is a blend of human and machine.
Why it matters: When fraud, hiring, and insurance underwriting all depend on knowing whether a human wrote something, "we can't reliably tell" is a systemic infrastructure problem, not just a content moderation headache.
Quick Hits
- →Anthropic's Fable 5.1 update cuts token costs and dials back overly aggressive content restrictions — a quiet but meaningful usability upgrade. TechCrunch
- →Mukesh Ambani's Jio wants to turn aging PCs into AI-ready machines for ~$11/month, targeting India's massive installed base of older hardware. TechCrunch
- →Google signed a deal to buy 400MW of geothermal power from Fervo — potentially expandable to 1GW — to feed its AI data centers in Utah. TechCrunch
- →Amazon's Alexa for Shopping can now verify whether a suspicious email or text actually came from Amazon — a direct shot at impersonation scams. The Verge
- →Google is partnering with MrBeast on a multi-year deal to feature Gemini, Google Health, and Fitbit in upcoming videos — starting with a wilderness survival challenge. The Verge
AI TLDR