OpenAI explains what actually caused the Codex file-deletion bug
Thibault Sottiaux laid out the root cause publicly: running with full access outside sandbox mode, the model tried to override the $HOME environment variable to set up a temporary directory, and — in what Sottiaux called an honest mistake — deleted the actual $HOME directory instead. This wasn’t a case of the model “going rogue” — it’s a plainer engineering failure. Once an agent has unsandboxed write access to your home directory, a single mistaken command is all it takes. Whether sandboxing should be on by default, and how much convenience developers are willing to trade away to keep it on, is a tradeoff every agentic coding tool now has to face.)
Kimi K3 and Inkling push open weights toward the 3T-parameter tier
Moonshot AI shipped Kimi K3, a 2.8T-parameter model it’s calling the first “open 3T-class” release, with weights promised by July 27. Around the same time, Mira Murati’s Thinking Machines Lab put out its first open-weight model, Inkling — a 975B MoE with 41B active parameters. Taken together, the open camp’s competitive axis has shifted from “will you open-source” to “how close can your open model’s parameter count get to the closed frontier” — a real stress test on training and inference cost. Whether the benchmarks back up the parameter-count narrative is the thing to watch next.
Anthropic makes Fable 5 permanent in Premium tiers
Starting July 20, Fable 5 becomes a standing model in Max/Team Premium subscriptions (Anthropic announced this only via its official X account — no anthropic.com newsroom post exists; the title links to that original post). The move itself isn’t surprising, but the timing — right on the heels of competitive pressure from GPT-5.6 Sol — suggests release cadence is increasingly being set by competitors’ shipping schedules rather than an independent roadmap.)
GPT-5.6 allegedly closed a 30-year gap in convex optimization — don’t buy it yet
A popular Reddit thread (also discussed on Hacker News) claims a single prompt to GPT-5.6 helped fill a proof gap that had stood open in convex optimization for 30 years, but the comment section is already picking at the rigor of the proof. This kind of “AI solves a decades-old math problem” story has been showing up on a near-monthly cadence, and the arc is familiar: a splashy headline first, followed by the math community slowly finding the holes. Until peer review actually confirms it, this belongs in the “worth watching” pile, not the “fact” pile.
Global Index on Responsible AI 2026 report released
GIRAI 2026 scores countries across five dimensions on how well their responsible-AI commitments actually get implemented. For anyone working in Responsible AI, cross-country quantitative comparisons like this are rare enough to be worth a closer look once the specific rankings and methodology are available.
One line for today: the Codex incident says the real constraint on agentic tools isn’t model capability — it’s permission design. Whoever ships “least privilege” as the default first is the one who avoids the next headline.