OpenAI is launching a personal finance feature for ChatGPT Pro users in the US, allowing bank account connections and surfacing a dashboard of spending, subscriptions, portfolio performance, and upcoming payments. This is OpenAI's first direct move into fintech, going head-to-head with Mint successors, Copilot, and newer AI-native PFM startups. The distribution advantage of ChatGPT's existing user base makes this an immediate threat to standalone personal finance apps.
The Musk v. Altman trial entered its final week with closing arguments focused on credibility: Altman was cross-examined on alleged self-dealing and dishonesty involving OpenAI partners, while Musk was painted as motivated by a desire to personally control AI development rather than altruistic safety concerns. The jury's verdict will have direct implications for OpenAI's nonprofit-to-for-profit conversion and governance structure. Regardless of outcome, the public testimony has already surfaced damaging narratives for both parties.
This newsletter edition covers two converging AI harm stories: non-consensual deepfake pornography affecting real individuals, and AI chatbots leaking real phone numbers from training or retrieval data. Both represent live, production-scale harms rather than hypothetical risks. The combination signals that AI safety regulators and plaintiffs' attorneys now have concrete, documentable cases to work from.
Google's generative AI search results are misdirecting callers to real individuals by hallucinating or incorrectly attributing phone numbers, causing documented harassment and business disruption to private citizens. The problem appears rooted in AI confidently surfacing stale, misattributed, or hallucinated contact data as factual. This is a production-scale harm with a clear litigation and regulatory surface area.
Simon Willison built and shipped a browser-based QR code generator — supporting both URL/text and WiFi network codes — using Claude as the primary development tool. The 290 HN score on what is functionally a simple utility signals strong community resonance with the 'vibe coding a useful tool in minutes' workflow. It's a data point, not a breakthrough, but it's a high-signal one about where developer productivity is landing in practice.
OpenAI's official announcement confirms the ChatGPT personal finance feature is live in preview for Pro users in the US, with secure financial account connections and AI-powered insights grounded in a user's actual financial context and goals. This is the primary source confirming what TechCrunch reported, with OpenAI emphasizing the security model and the contextual, goal-aware nature of the AI guidance. The 'grounded in your financial context' framing suggests this is RAG over live account data, not just generic financial advice.
Cerebras went public in 2026 raising $5.5B and saw its stock more than double on debut, signaling strong public market appetite for AI infrastructure plays beyond Nvidia. The IPO validates the wafer-scale chip architecture as a legitimate alternative compute path and opens the capital markets for other AI infrastructure companies. This is the first major tech IPO of 2026 and will likely serve as a sentiment bellwether for the broader AI infrastructure funding cycle.
A second severe Linux kernel vulnerability has been disclosed within two weeks, with production patches now available. The frequency of critical Linux CVEs is increasing, creating patching fatigue and compressing response windows for teams running Linux-based AI infrastructure. Unpatched Linux hosts remain the most common attack vector for cloud infrastructure compromise.
Richard Socher's new startup has raised $650M to build a self-improving AI system capable of conducting its own research and iterating on its own architecture indefinitely. The company is explicitly positioning against the 'research lab without products' trap, claiming it will ship commercial products alongside the research agenda. Self-improving AI is one of the most capital-intensive and uncertain bets in the space, but $650M buys significant runway to find out.
Carnegie Mellon researchers built a benchmark demonstrating that Claude Mythos and GPT-5.5 can autonomously develop working exploits for real vulnerabilities in Google's V8 JavaScript engine. Mythos leads GPT-5.5 significantly on the benchmark but costs 12x more per task. This is a qualitative capability threshold: AI systems are now crossing from 'assist a hacker' to 'autonomous offensive security research.'
That's today's briefing.
Get it in your inbox every morning — free.
Help us improve AI in News
Got a suggestion, bug report, or question?