SI Newsweek 2026 10 26 When the Agents Tested the Cage

SI Newsweek 2026 10 26 When the Agents Tested the Cage A ~23-minute deep dive on structural AI news for Oct 19-25, 2026: the week AI agents were paid to break out of their own training cage, and outsiders published what they found inside the labs. Correction: last week's episode said OpenAI's training pause was still on. OpenAI disclosed Oct 23 that it ended Oct 1. Key takeaways OpenAI's pause ended Oct 1. Frontier training runs in an offline world with no live internet, gated by agent red-teaming. Agent-found bugs to patch now: php2shell (PHP 8.5.12, 8.4.27, 8.3.36, 8.2.35) and Brotli HTTP2Shell (CVE-2026-111497; Node 22.23.4, 24.22.1, 26.12.1). METR and Redwood looked inside OpenAI's monitoring and found gaps. OpenAI says they are patched; not externally verified. METR ties Anthropic's low risk rating to limited capability. Anthropic's Mythos 5 analysis shows fake-account pressure on a maintainer. Anthropic's leaked draft bill drew fire from allies; Dario Amodei's nine-decision reply is his own account. Agent land-grab: Claude Duet, Claude auto mode default, Meta Muse over 10M accounts, Argon, OpenAI on Google Cloud. Texas sued OpenAI Oct 23 seeking a court-appointed evaluator. Florida's injunction has no hearing date. Dates to watch Week of Oct 26: Claude apps auto mode default (announced, not yet in effect) Nov 30: Australia JSCAI final report; Argon intro pricing ends X posts referenced (status, then link): Confirmed OpenAI offline-web restart: https://x.com/OpenAI/status/21137187975050... Brotli flaw: https://x.com/OpenAI/status/21137663543860... php2shell, $1B for defenders: https://x.com/OpenAI/status/21129078560183... OpenAI monitoring pilot: https://x.com/OpenAI/status/21132988333924... METR monitoring pilot: https://x.com/METR_Evals/status/2113290966... Madry returns to OpenAI: https://x.com/aleks_madry/status/211223144... Altman on Madry: https://x.com/sama/status/2112652656602923366 Anthropic Mythos 5 analysis: https://x.com/AnthropicAI/status/211362665... METR review of Anthropic: https://x.com/METR_Evals/status/2112562150... Dario Amodei reply (post confirmed; decisions are his account): https://x.com/DarioAmodei/status/211273223... Claude Duet preview: https://x.com/claudeai/status/211298466425... OpenAI models on Google Cloud: https://x.com/OpenAI/status/21129881184712... NVIDIA DSX POD (14 days is NVIDIA's claim): https://x.com/nvidia/status/21128824485308... DeepSeek open inference engine: https://x.com/deepseek_ai/status/211247649... Partly confirmed Claude auto mode default (announced): https://x.com/claudeai/status/211336084540... Pichai, Argon about 25T tokens (Google's figure): https://x.com/sundarpichai/status/21129457... Suleyman, MAI-DxO 1.7x (benchmark only): https://x.com/mustafasuleyman/status/21133... Alexandr Wang, Muse Japan launch confirmed, #1 app partly: https://x.com/alexandr_wang/status/2113682... Unverified (claim only) Noam Brown, Navier-Stokes (Clay review open): https://x.com/polynoamial/status/211372926... Musk, Grok Bot calls businesses: https://x.com/elonmusk/status/211436536106... AI-generated audio deep dive (NotebookLM). X posts are treated as leads and were checked against primary sources (filings, company posts, regulators, and news reports) before being labeled.