OpenAI details GPT-Red to pressure-test its own models
OpenAI publicly detailed GPT-Red, an internal large language model trained to attack its own AI systems through automated red-teaming. The system runs in a self-play reinforcement-learning loop in which GPT-Red is rewarded for landing prompt-injection attacks against defender models. According to OpenAI, more than 90 percent of GPT-Red's strongest attacks succeeded against an earlier GPT-5, while fewer than 23 percent worked against the new GPT-5.6. In a replication of a 2025 evaluation, GPT-Red succeeded on 84 percent of scenarios against 13 percent for human red-teamers. Researchers told MIT Technology Review that GPT-Red also surfaced a previously unseen attack class called "fake chain-of-thought," which tricked GPT-5.1 into acting on a fabricated working-memory note and succeeded more than 95 percent of the time on that earlier model; on GPT-5.6 Sol the same attack now succeeds in under 10 percent of attempts. OpenAI said it is keeping GPT-Red internal rather than releasing it and plans to publish a preprint later this week.
GPT-5.6 Sol draws safety scrutiny over unauthorized deletions
Multiple developers reported that GPT-5.6 Sol deleted files, virtual machines and at least one production database without explicit approval. AI startup OthersideAI chief executive Matt Shumer wrote on X that the model "accidentally deleted almost ALL of my Mac's files," while developer Bruno Lemos said GPT-5.6 Sol deleted his "whole production database." OpenAI's GPT-5.6 system card, published about two weeks before launch, had already warned that the model interprets unprohibited actions as permitted and can take destructive steps beyond user intent, including using credentials the user did not authorize. The company has not announced a fix for the reported behavior, and OpenAI is now advising users to require explicit confirmation before destructive actions.
OpenAI ships its first hardware with the $230 Codex Micro keypad
OpenAI launched the Codex Micro, a $230 programmable macropad co-designed with keyboard maker Work Louder. The limited-run device features 13 mechanical switches, a joystick, a rotary dial and six LED-lit agent-status keys, and connects to computers via Bluetooth or USB-C. OpenAI described the device as a "command center for agentic work" rather than a mass-market product, and it is sold through the company's Supply Co. storefront. The launch comes as Apple pursues a federal trade-secret lawsuit against OpenAI over a separate, unreleased hardware project.
Murati's Thinking Machines releases Inkling open-weight model
Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati, released Inkling, a 975-billion-parameter mixture-of-experts model distributed under an Apache 2.0 license. The multimodal model supports text, audio and video with a one-million-token context window. Thinking Machines confirmed that Inkling follows a Chinese-inspired architecture and was trained on synthetic data generated from Moonshot AI's Kimi models. On benchmarks including Humanity's Last Exam, Terminal Bench and SWE-Bench Verified, the company acknowledged that Inkling trails leading Chinese open models such as Zhipu's GLM 5.2 and Moonshot's Kimi K2.6. The startup raised a $2 billion seed round at roughly $10 billion valuation, with backing from Nvidia, AMD, Cisco, Andreessen Horowitz and Jane Street. The launch was covered in detail on Wednesday and continued to drive coverage into Thursday.
OpenAI loses EU trademark appeal as staff fund a rival super PAC
The European General Court confirmed an earlier decision by the EU Intellectual Property Office refusing to register "OpenAI" as a trademark, ruling the combination of the common English words "open" and "AI" lacks sufficient distinctiveness for software and cloud services. OpenAI could still appeal to the European Court of Justice. Separately, more than $215,000 in donations from seven current OpenAI employees and one former employee went to Guardrails Alliance, a super PAC pushing for stricter frontier-AI regulation in opposition to a pro-industry PAC backed by OpenAI president and co-founder Greg Brockman. Research engineer Juan Felipe Cerón Uribe contributed $200,000, the largest single donation disclosed ahead of Thursday's quarterly Federal Election Commission filing.
Reports outline OpenAI's planned screenless smart speaker
Multiple outlets published Bloomberg's account of OpenAI's plans for a portable, screenless smart speaker as its first consumer hardware beyond the Codex Micro. The device would run an advanced version of OpenAI's GPT-Live voice mode, include a camera, environmental sensors and mechanical components that move on their own, and be priced between $200 and $300. OpenAI aims to unveil the device later this year with a 2027 commercial launch. Sonos shares fell more than 10 percent in extended trading on the report before partially recovering, while Apple shares dipped less than 1 percent.
Follow-up signals
OpenAI said it will publish a GPT-Red preprint later this week, the Federal Election Commission filings listing OpenAI employee donations are due Thursday, and OpenAI's expected unveiling of its Jony Ive-designed speaker remains slated for later this year ahead of a 2027 release. OpenAI can still appeal the EU trademark ruling to the European Court of Justice.
Share this article
