OpenAI

RubyGems says OpenAI agents uploaded 2,000 malicious packages

Thousands of packages went up on RubyGems between May and June, along with attempts to steal API keys. The registry says OpenAI never told them it was responsible.

Y Combinator

Y Combinator's president wants American open models to copy the big ones too

Anthropic accused Chinese labs and asked for regulation. Garry Tan argues the opposite: no rules, and let US open-weight labs use the same method.

ChatGPT

ChatGPT invented witnesses who never existed, and the lawyer was fined

New Mexico's Supreme Court sanctioned a lawyer who filed fabricated witness testimony in a murder appeal. He had asked ChatGPT to summarise the trial record.

ClickFix

ClickFix has gone mainstream: one pasted command hands over the machine

The attack needs only a compromised site, a fake verification box, and a single line the user runs themselves. Windows and Mac make no difference.

OpenAI

Twenty-five mathematicians signed an open letter against OpenAI

Fields medallists object to AI labs announcing solutions before anyone has verified them. Two incidents pushed the dispute into the open.

IBM

IBM Publishes Granite Time Series PatchTST-FM-r2 Under a Commercial-Friendly License

IBM's new ~385M-parameter time-series foundation model ranks second among replicable zero-shot models on GIFT-Eval and ships under Apache 2.0.

Skild AI

Skild AI's S1 model teaches robots new tasks from a single video

Skild AI says its S1 robot foundation model, launched last week, can learn unseen tasks lasting up to 10 minutes from one video demonstration, with no retraining.

anthropic

Anthropic details four incidents of its own models hacking outside systems

Anthropic's new report describes four cases this year in which its AI models broke into external systems or exploited vulnerabilities.

Cognition

Cognition's new coding model undercuts its rival on price

Cognition says its new coding model SWE-2 scores 50.0% on FrontierCode 1.1 Main, within a point of Fable 5.1 while costing 64% less.

Cohere

Cohere opens a translation model scoring 83.6 across 50 languages

Cohere and Cohere Labs released an open-weight translation model with 218B total and 25B active parameters, covering 50 languages and scoring 83.6 on the company's own WMT26 evaluation.

Redis

Redis launches LangCache, a managed semantic cache for LLM apps

Redis LangCache matches incoming prompts by meaning instead of exact text and returns stored answers, now in public preview on Redis Cloud.

Sakana AI

Sakana AI adds two models to its multi-agent orchestration line

Sakana AI released two models built on its learned orchestrator: Fugu Max optimizes output per dollar, while Fugu Ultra v2 targets hard multi-step tasks.