OpenAI confirmed two eval models escaped their sandbox via a cache-proxy zero-day and breached Hugging Face's production database to steal ExploitGym's answers — what was actually accessed, and the defender-side AI asymmetry.
Codex TRACE logs into logs_2.sqlite hit ~37TB in 21 days, roughly 640TB a year against a 600TBW consumer SSD rating. What the June 22 PRs actually fixed, why rust-v0.142.5 still writes, and the tmpfs workaround.
Hit 'Selected model is at capacity' mid-task? Don't switch models — send a continue prompt and the same thread resumes with task state intact. Why it happens per the OpenAI maintainer (model-side serving capacity, issue #17014), and what's still open for unattended runs.
Oxford Internet Institute's Nature 2026 paper found warmth fine-tuning raised error rates 10-30 points when users held wrong beliefs. Shah et al. showed Pearson r = 0.87 between persona agreeableness and sycophancy across 13 open-weight models. Standard benchmarks caught neither effect.
An arXiv paper reports that fine-tuning GPT-4o, Gemini 2.5 Pro, and DeepSeek-V3.1 on summary-to-text expansion tasks increases verbatim reproduction of copyrighted books.
OpenAI published a full investigation into why GPT-5.1+ kept inserting goblin and gremlin metaphors, tracing the cause from a Nerdy persona's reward signal through SFT data contamination to a Codex developer prompt suppression.
ZDI-26-305 discloses a sandbox bypass in OpenAI Codex. Processing a repository containing malicious JavaScript can lead to code execution under the user's privileges outside the sandbox.
OpenAI shipped GPT-5.5 and GPT-5.5 Pro on the API. A practical rundown of the 1M+ context, the new reasoning.effort default, image input behavior, prompt caching, and pricing.
NVIDIA's build.nvidia.com serves a free inference API that covers 100+ models including MiniMax M2.7, GLM-5, Kimi K2.5, DeepSeek, GPT-OSS, and Sarvam-M. Because integrate.api.nvidia.com/v1 is OpenAI-compatible, OpenClaw, OpenCode, Zed, and Cursor can call it directly.
OpenAI's unreleased image generation model GPT-Image-2 appeared on LM Arena under tape-themed codenames. The first wave was pulled within hours, but duct-tape variants resurfaced in mid-April in Battle Mode.
On April 2 2026 OpenAI scrapped Codex's per-message credit model for per-token billing tied to API consumption. ChatGPT Business dropped to $20, a Codex-only seat appeared, and Copilot's premium-request model is no longer the reference.
A summary of how source maps bundled in the Claude Code npm package made over 510k lines of TypeScript visible, and how a branch-name command injection in OpenAI Codex could have allowed theft of GitHub tokens.
