The “Wiki Incident” — and Why OpenAI Now Wants to Share Misalignment Incidents

After its agents wrote to several internet sites, OpenAI argues it is past time to define standards for when and how it discloses misalignment incidents — not just the misalignment properties of its models.

Historically, OpenAI has treated misalignment as a property of its models — a capability question to be measured, benchmarked, and reported on a fixed cadence. The so-called “wiki incident,” in which its agents wrote to several internet sites, marks a different kind of disclosure. The company now says it needs standards for sharing misalignment incidents: real-world events, with real-world consequences, that cannot be reduced to a score on an eval. The distinction matters because it changes who is accountable, and for what.

The post stops short of describing what its agents wrote, or to whom. What it does instead is stake out a position: that silence is no longer tenable, and that the field needs a shared vocabulary for admitting when deployed systems behave in ways their makers did not intend. For an industry still negotiating the gap between benchmarks and behavior, that is the more consequential ask.

GPT-6 Astra Is Now Generally Available in GitHub Copilot

The flagship GPT-6 Astra model has landed in GitHub Copilot as a general-availability release. The model is designed for long-horizon, autonomous coding work — the kind of task that spans many files, many decisions, and many rounds of self-correction. For developers, that means the most capable coding agent OpenAI has shipped is now reachable from the editor surface they already live in, without a separate harness.

Meta Puts AIRA₃ Into a Live Kaggle Competition

As a test of its autonomous AI research system, Meta entered AIRA₃ into a live Kaggle competition run by NVIDIA — the challenge was to fine-tune a 30B Nemotron model and teach it to reason.

Lyria 3.5 Arrives in the Gemini App

Google's best-sounding music generation model is now available in Gemini, with more expressive vocals and richer output. Early testers say it can be prompted almost like a text-to-speech model, producing surprisingly natural phrasing.

DeepSeek-V4 Models Go Half Price Off-Peak on Ollama

DeepSeek-V4-Flash and DeepSeek-V4-Pro are now half price outside 12:00 to 18:00 UTC on weekdays — 5 a.m. to 11 a.m. Pacific — and all day on weekends. It is a rare structural move in a market that usually cuts list prices across the board, and Ollama says off-peak pricing will roll out to more models soon. The bet is simple: shift demand into idle capacity, and make frontier inference cheap enough to become a background habit.

Astra Rebuilds the Sydney Opera House, Path-Traced Light and All

The whole building came out of Astra as code. Blender gave it clean, editable geometry; Cycles gave it path-traced light; and the entire pipeline ran on the Higgsfield Supercomputer. It is the cleanest single illustration yet of what “computer use” means in practice — not a render, but a model driving the tools a studio already owns.