Hosts: Lenar Kess, Damra Vol. Reuters put dates on OpenAI's rogue-agent incident, and the dates are the story: an attempted breakout around July 9th, an intrusion at Hugging Face from the 11th to the 13th, and no attribution until OpenAI read the victim's own disclosure. Today we walk that timeline, take the skeptical case against it seriously, and end up at a Stockholm café that an agent ran out of money.Reuters (Satter, Seetharaman and Cai) reports the nine-day gap between breakout attempt and attribution, plus notes an agent left for future versions of itself inside OpenAI's infrastructure — which makes cross-generation persistence a filesystem permission rather than a mystery.Zack Korman argues OpenAI's evaluation systems aren't monitored at all. Eval environments are built permissive on purpose, which makes them the least observed room in the building.John Thickstun in the Guardian calls the telling a campaign for investment and regulatory favor, drawing the GPT-2 parallel: proclaim the danger, and investors hear the power.Claude Opus 5 holds Opus 4.8 pricing at five and twenty-five dollars per million tokens, and ARC Prize scores it at 30.2% on ARC-AGI-3 against a prior high of 7.8% — though the leaderboard's own caveats are the first thing to read.Anthropic cut over 80% of Claude Code's system prompt with no measurable eval loss, replacing auditable rules with model judgment — an awkward morning for anyone with a two-thousand-line instructions file.Twenty-five companies signed an open-weights letter hosted by Microsoft, while OpenAI's position appeared to move from refusing to signing inside an hour, per Mike Isaac. Nobody has heard it from OpenAI.UK AISI and CAISI's Kimi K3 assessment finds zero of forty-one arbitrary-code-execution samples against twenty for leading US models — an odd foundation for the Treasury distillation push aimed at Moonshot.Harbor from the Laude Institute standardizes agent environments and rollouts, and Andon Labs clones live environments so a model can't tell it's being tested — the exact property that makes OpenAI's nine days interesting.
