
OpenAI Built GPT-Red to Attack Its Own Models Before Hackers Do
OpenAI built GPT-Red, an internal automated red teamer that attacks AI agents, finds prompt-injection vulnerabilities, and helped harden GPT-5.6 Sol.
Search the archive
Archive
A tag for posts related to OpenAI, including news, tutorials, and discussions.

OpenAI built GPT-Red, an internal automated red teamer that attacks AI agents, finds prompt-injection vulnerabilities, and helped harden GPT-5.6 Sol.
Anthropic extends Fable 5's deadline while OpenAI ships GPT-5.6 Sol in every subscription. The contrast says everything about who respects their users.

ChatGPT Work turns a clear goal, your files, connected tools, and approved actions into finished documents, spreadsheets, presentations, analyses, workflows, and Sites. This complete…

ChatGPT Skills turn your best prompts, processes, examples, and quality rules into reusable workflows. This complete guide covers how to find, install, invoke, create,…
Matt Shumer says OpenAI's GPT-5.6 Sol in Ultra mode ran rm -rf on his dev directory during a stress test. A concrete AI safety…

Apple filed a federal lawsuit against OpenAI on Friday, accusing the AI company of systematically stealing trade secrets to build a competing hardware device.…

A community Hermes Agent skill merges GPT Image 2, Gemini Omni Flash, and Suno into a single pipeline. The skill works great. The skill…

OpenAI is shutting down its AI-powered browser Atlas after nine months. But the Chrome extension and cloud desktop browser replacing it represent a bigger…

OpenAI's No. 2 executive is leaving her full-time role after a prolonged medical leave, leaving Sam Altman without a key lieutenant at a critical…

OpenAI launched ChatGPT Work, a multi-hour agent that runs across apps and files. It's powered by Codex, GPT-5.6, and lands two days after Claude…

OpenAI launched the GPT-5.6 family to general availability on July 9 with Sam Altman calling it their best model ever. Sol, Terra, Luna, and…

OpenAI audited SWE-Bench Pro and estimates about 30% of its tasks are broken. The company is retracting its recommendation that the research community use…