An internal document obtained by Wired reveals Meta secretly ran a project codenamed 'Cannes' where outsourced workers from Covalen in Dublin created fake underage accounts to bombard ChatGPT, Gemini, and Character.AI with disturbing prompts involving self-harm, suicide, eating disorders, and sexual content. In one 2025 testing round alone, over 45,000 high-risk prompts were submitted. The targeted companies were not informed; Character.AI stated the actions violated its terms, while OpenAI and Google said they had not authorized such testing. Meta defended the work as standard safety benchmarking, but external experts called it a case of anti-competitive behavior using AI safety as a cover. Multiple workers reported psychological distress from the content they were forced to generate.
A roundup of July 2 tech news: Apple is reportedly planning a major Apple Watch redesign for 2027 with a new band connector, making existing bands incompatible. Anthropic's export restrictions on Claude Fable 5 were revoked, and Claude Code's hidden detection mechanism for Chinese users will be removed. A MacBook Ultra with OLED touchscreen and M5 Pro/Max chips is rumored for late 2026–2027. Xiaomi's first NAS crowdfunding sold out within an hour. SiliconFlow filed for IPO in Hong Kong, aiming to be the 'first AI token factory stock.' Baidu hired 97-born AI expert Sun Tianxiang to lead its foundational model R&D unit. Google released Nano Banana 2 Lite, a fast and cost-efficient image generation model. Xpeng's MONA L03 SUV will debut on July 2 with 1500 TOPS computing power and VLA. Samsung Galaxy Glasses design was revealed via a companion app. Microsoft plans new layoffs targeting sales and consulting roles. Panasonic announced a ~500 billion yen investment shift to AI infrastructure.
Anthropic released Claude Sonnet 5, its most agentic Sonnet model with improved reasoning, tool use, and coding. Entry API pricing is $2/million input tokens and $10/million output tokens until August 31, 2026, then reverting to $3 and $15, while a new tokenizer may increase token consumption by up to 35%. The model is default for Free and Pro users and available via API and Claude Code. Simultaneously, Anthropic launched Claude Science beta, a scientific workbench integrating over 60 skills and databases for reproducible research and on-premises data processing, with a project grant program. The announcements follow Anthropic's IPO filing and reported $47 billion revenue run rate.
Cursor launched an iOS app that turns the iPhone into a remote control for its AI coding agent, enabling users to start cloud agents, receive push notifications, review diffs and screenshots, merge PRs, and use voice commands. OpenClaw released native iOS and Android apps that extend its personal AI agent to mobile, allowing the agent to access the phone’s camera, microphone, location, notifications, and push-to-talk for approvals and task execution. Both apps reflect a shift from the chatbot paradigm toward agents working continuously in the background while users intervene only at key decision points via mobile. Cursor’s app requires a Pro subscription and is iOS-only; OpenClaw’s apps require pairing with the desktop version. The trend points to the smartphone becoming a supervisory console, freeing people from having to sit in front of a computer while AI does the heavy lifting.
A ransomware leak from Tata Electronics exposed hundreds of GB of iPhone 18 Pro files, including A20 chip and supplier details. SK Hynix filed for a Nasdaq IPO under ticker SKHY, aiming to raise capital as the HBM market leader. Anthropic launched Claude Sonnet 5, a mid-size model focused on lower-cost AI agents with improved tool use and safety, priced at $2/$10 per million input/output tokens. ByteDance CEO Liang Rubo issued new leadership principles demanding hands-on output from managers. Kimi's latest funding round raised its pre-money valuation to $31.5 billion with an ARR exceeding $300 million. Other highlights include Ubtech's $16,980-$99,000 U1 humanoid companion robot, Huawei's open-source 92B MoE model openPangu-2.0-Flash, Meituan's VitaBench 2.0 for long-term agent user modeling, and research warning that labeling AI as “employees” reduces human error detection by 18%.
OpenAI announced the Codex Micro, a co-branded programmable macro pad developed with peripherals maker Work Louder, at the AI Engineer World’s Fair. The device closely resembles Work Louder’s Creator Micro 2 and features 13 mechanical keys, a joystick, and a touch sensor. It is designed to work with Codex, OpenAI’s AI coding agent with over 5 million weekly active users, enabling developers to trigger code completions, error corrections, and context switching without leaving their editor. The hardware serves as a physical productivity anchor for desktop workflows, while OpenAI separately pursues a consumer AI device with former Apple designer Jony Ive.