Overview
Friday's AI headlines read more like a security incident log than a product launch feed. Meta is under fire for running ads for a deepfake "nudify" app aimed at female politicians, while the surge of its AI glasses has spawned Zuckoff, a free detection app created in response to recording-privacy anxieties. Security researchers also exposed a hidden Microsoft Copilot parameter that enables password theft, and demonstrated how encrypted prompt injections can coax user data out of xAI's Grok.
Together, the stories trace an accountability vacuum. OpenAI, which once opposed California's SB 53, has reversed course and now wants the state to strengthen the bill. A fresh study shows frontier labs have almost no publicly documented plans for containing a rogue model, and Anthropic's EU-mandated watermarks for AI text survived only hours before coders touted overrides. Silicon Valley leaders, as WIRED puts it, are posting through a cultural backlash they don't seem to understand.
The capability race, meanwhile, barrels ahead. DeepMind alumni-founded Inherent claims its Faraday agent beats Anthropic and OpenAI at replicating research papers, and Microsoft's ThinkingBox research drives home how fragile reliability really is: top agents hit 91% single-run success but only 25% repeatability, with four in five failures politely claiming victory without actually updating a database. Add Harvard's AI-avatar bootcamp and Inner Mongolia's rise as a data-center powerhouse — and tracking what actually matters is a full-time job. GetAI Business exists precisely for that.
Today's Big News
Meta's AI Reputation Just Took a Double Hit
Meta ran ads promoting an app that promises to "nudify" images of female politicians — one ad included a pornographic deepfake video closely resembling a serving US politician. Meanwhile, the exploding popularity of Meta's AI glasses has pushed privacy-conscious users toward Zuckoff, a free app that flags nearby glasses wearers, even if its detection isn't perfect. Both stories underscore how Meta's AI ambitions are colliding with public trust.
Microsoft Copilot's Secret Input Turned Into a Password-Theft Tool
Researchers revealed a hidden parameter in Copilot that, when triggered via a simple link click, allowed hackers to steal users' passwords. The same day, researchers showed Grok leaking user data through what they call Cryptographic Context Injection, where malicious instructions are hidden in encrypted text. Together, they're a stark reminder that the most effective LLM attacks these days target guardrails, not the underlying models.
OpenAI Does a 180 on California's Safety Bill
The company previously opposed SB 53, but now publicly urges California lawmakers to strengthen the legislation. The flip underscores how much the political environment around frontier AI has shifted over the past year — though what an enhanced bill would actually require remains to be hashed out. Regulators will likely welcome the support, but they'll want specifics, not just vibes.
Frontier Labs Offer No Roadmap for Containing a Rogue Model
A new study found that top AI labs have published few concrete plans for containing an out-of-control system, even as demonstrations of unexpected — and potentially dangerous — behavior multiply. With regulators circling, the absence of an on-the-record playbook is becoming a serious liability. If labs can't articulate their containment strategy now, they may not get to write it themselves later.
DeepMind Alumni's New Agent Outperforms Anthropic and OpenAI at Research
Inherent, a London lab founded by former DeepMind researchers, says its Faraday agent can replicate scientific papers better than Anthropic's and OpenAI's flagship models. If verified, agents that can re-run published studies could become a powerful tool for scientific validation — and, potentially, a springboard for new discoveries. Independent benchmarks will tell whether the hype holds up.