AI Models Found a Web-Wide Bug

OpenAI says its models helped find critical flaws in PHP, the code behind an estimated 70% of websites.

Cloudflare warns the PHP bug and a separate certificate flaw mean everyone should upgrade their OS, browser, and servers now.

Anthropic opened a research preview of Duet, an agent that works while you step away and checks before acting.

Greg Brockman says the PHP bug sat undetected for 14 years before GPT-5.6 Sol helped uncover the full version in August.

Stripe is wiring into Duet so people can make purchases and issue refunds straight from the agent.

A tester says Duet is the first personal agent worth calling a pro tool, shifting half his usage from ChatGPT to Claude.

Anthropic's new Sonnet 5.5 model scored 72.4% on the hardest AI reasoning test, ARC-AGI-3.

Legal AI startup Legora raised a $550 million round at an $8.8 billion valuation, going from $1 million to $100 million ARR in 18 months.

Remote's Alex Bouaziz says the best AI models only solve 49% of real payroll tasks correctly, prompting a new AI lab to fix it.

GitHub previewed a tool letting teams ship AI-written code to production without reviewing it line by line.