🤖 OpenAI ships GPT-Red to stress-test GPT-5.6 Sol 🤖
Automated red-teaming speeds up safety testing for OpenAI's latest production model.

OpenAI has published research on GPT-Red, a purpose-built model trained to iteratively generate adversarial prompts and uncover vulnerabilities in other AI systems at scale. By feeding GPT-Red's attacks back into the training pipeline, OpenAI reports a sixfold reduction in failures on a challenging prompt-injection benchmark for its latest GPT-5.6 Sol model. The approach — using AI to stress-test AI — could standardize how the industry scales safety testing. Unlike manual red-teaming, which struggles to keep pace with increasingly capable models, GPT-Red can probe for failure modes continuously and at far greater volume. The research underscores OpenAI's push to harden production models through adversarial training loops before public deployment.

🧠 Thinking Machines drops Inkling 975B model 🧠
A 975B-parameter open-weights model rolls out with an adjustable reasoning effort dial.

Thinking Machines has released Inkling, a 975B-parameter mixture-of-experts model with 41B active parameters, multimodal reasoning, and a one-million-token context window. Rather than chasing raw benchmark scores, Inkling focuses on customizability: an "effort dial" lets users trade thinking depth for cost, and on one coding test it matched Nvidia's Nemotron score using just one-third the tokens. The open-weights model is available on Hugging Face and can be fine-tuned through Thinking Machines' Tinker training service. While it still trails top Chinese open models on some benchmarks, it adds real competitive pressure to the open-weights ecosystem, showing that smaller labs can build large-scale models through smart architectural choices rather than brute compute.

🔓 SpaceXAI open-sources Grok Build amid data controversy 🔓
Grok Build's rollout is overshadowed by a data-handling discovery that shook developer trust.

SpaceXAI open-sourced Grok Build, a terminal-based coding agent that can inspect codebases, edit files, run shell commands, search the web, and manage long-running tasks. The tool supports interactive use, headless scripting, CI pipelines, and editor integration through the Agent Client Protocol. However, the rollout was quickly overshadowed: a security researcher discovered the CLI had been shipping entire user repositories to an xAI-controlled Google Cloud storage bucket. SpaceXAI responded by pledging to delete all retained coding data, though it denied using it for model training. The incident highlights the trust gap in AI developer tooling — a critical issue as coding agents gain broader adoption. The story drew coverage from five sources, reflecting significant developer community concern.

INTERESTING TO KNOW

⌨️ OpenAI ships $230 Codex Micro keypad ⌨️

OpenAI's first hardware rollout is a $230 mechanical keypad called Codex Micro, built with Work Louder to give developers tactile control over AI coding agents. The device features color-coded "Agent Keys" showing real-time task status, a joystick for toggling between jobs, and a dial to adjust reasoning depth — available in both "Clicky" and "Silent" switch variants. Sold through OpenAI's "Supply Co." merch section (not part of the rumored Jony Ive initiative), it signals a new hardware distribution channel for AI workflows.

🔬 Weco's AIDE² achieves recursive self-improvement 🔬

AI research team Weco has demonstrated the speed at which AI agents can now optimize themselves: their system AIDE² redesigned its own research process over an eight-day run, outperforming a version that human engineers refined over two years. Operating through two connected cycles — one agent solving research problems, another revising the first's strategy — AIDE² tested 100 rewrites, retained 7 meaningful upgrades, reduced prompt size by 16x, and cut reward-hacking rates from 63% to 34%. This raises urgent questions about autonomous capability jumps in self-improving systems.