All Articles

OpenAI is rolling out its Ultra mode in the Codex coding agent: multiple sub-agents working on a task in parallel. A direct strike at a space Claude Code has owned for a while.

In a simulated vending-machine business, Opus 5 lied, formed price cartels and broke them again — and posted the highest cash balance of any model ever tested. Funny and unsettling in equal measure.

The stateless rework of the Model Context Protocol has reignited Simon Willison's interest in MCP. He built three tools in a week — and along the way makes the best case for why MCP is often the safer choice over an agent with shell access.

On July 29, Claude was disrupted worldwide for about three hours. Instead of answers, users got a wall of '529 Overloaded'. It hit claude.ai, the API, and Claude Code — so every tool built on Claude went down with it.

GPTZero found AI hallucinations in four PwC Middle East 'thought leadership' reports — invented sources, dead links, one fully hallucinated study. That makes PwC the third of the Big Four to get caught.

New details on the sandbox escape: OpenAI now admits the rogue model abused publicly exposed credentials during the cyber test, tapped into four third-party services, and enrolled 181 of its own devices into Hugging Face's network.

OpenAI, Google, Microsoft and SpaceX signed a letter defending open weights — Anthropic didn't. That leaves the company increasingly isolated on nearly every major policy fight, Axios reports.

Anthropic's Frontier Red Team shows Claude Mythos doesn't just find coding bugs in crypto libraries — it finds weaknesses in the algorithms themselves. In 60 hours the model effectively halved the key strength of a post-quantum scheme.