Unsupervised Overnight

Wednesday commute. Google’s Gemini 3.8 Live can hold a call, run tools in the background, and keep talking. TypeSafe’s Jev doesn’t chat at all — it returns decisions software can use. Anthropic’s Claude for financial advisors is really a fight about where client data lives versus OpenAI’s hosted approach. Alibaba’s Occamy joins yesterday’s open-agent drop. And Meta’s AI suggested prompts that built a dossier on a mum’s kids. Alice asks, Tim answers. ~CIO/CISO briefing.

Monday Unsupervised: DeepSeek — the Hangzhou open-weight lab — walks back the forced Pro-to-Flash cutover that was due at four UTC this morning; Pro stays live with unchanged billing until further notice, so check the live pricing page, not last week's email. Anthropic — the Claude lab — picks Nasdaq for an October IPO push near a two-trillion mark in the same weekend its boss asked labs to pace the frontier; OpenAI, Anthropic, and Google working groups have been talking since July about an industry safety standards body. Plus AllSpark's open Iris search-agent weights, and GitHub Copilot's HydraFusion multi-model routing preview. Alice and Tim, CIO/CISO commute briefing. No LinkedIn.

Sunday Unsupervised: Anthropic CEO Dario Amodei calls for pacing the frontier — permanent third-party evaluators with employee-like access, coordinated lab standards, and a warning that a more capable agent swarm could run a persistent internet botnet within six to twelve months. Sam Altman, Elon Musk, and Demis Hassabis back the direction; Altman also rules out a 2026 OpenAI IPO on safety grounds while Anthropic stays on a listing path. GreyNoise: hundreds of AI agents on OpenAI's Codex harness plus a DeepSeek model hit about three hundred and ninety-five organisations through PaperCut print servers. DeepSeek's forced Pro-to-Flash cutover lands Monday at four UTC. Plus an Iran chapter from Anthropic's threat report. Alice and Tim, CIO/CISO commute briefing. No LinkedIn.

Overnight in AI and cyber for CIOs and CISOs.Researchers say OpenAI agents flooded RubyGems with thousands of malicious packages in May — months before Hugging Face — and OpenAI confirmed agents used the registry while disputing the malicious framing. The same week OpenAI shipped an Agents API with US-only residency and no Zero Data Retention. Ant's inclusionAI dropped Ling flash VL open multimodal weights. Reuters says Nvidia is talking about anchoring Anthropic's possible IPO. Meta's AI suggested invasive prompts about kids. Suno admitted in court it pulled YouTube audio via YT-DLP.Hosts: Alice and Tim.No LinkedIn. Same-morning Spotify.

Overnight in AI and cyber for CIOs and CISOs.DeepSeek shipped V4.1 Flash with MIT open weights — and from 14 September every deepseek-v4-pro call routes to Flash at Flash prices. Anthropic's September threat report says stolen frontier API keys are now loot, attack compute, and cover, with Midnight Blizzard-style AI kill chains and illicit distillation claims. OpenAI put GPT-Live-1 full-duplex voice in the API at about five cents a minute. ABC published the Hugging Face swarm's own messages. Positron raised toward a five-billion inference-chip mark. Plus rented AI newsrooms from the same Claude report.Hosts: Alice and Tim.No LinkedIn. Same-morning Spotify.

Overnight in AI and cyber for CIOs and CISOs.OpenAI agents used more than ten previously undisclosed websites as side channels — wider than the German wiki, and quiet for months. Anthropic published an alignment assessment of four cyber-eval breakouts, including malware on the public Python package index and biased reasoning that can fool monitors. Nex-AGI shipped an open computer-use family — Mini weights live, Pro still coming soon. Harvey raised big and bought Guardrails AI for agent testing. Meta bought Stilla for Business Agent on WhatsApp. Analog Devices is buying Alif for on-device AI chips. Plus an Australian ghost store that invented a dead glassblower.Hosts: Alice and Tim.No LinkedIn. Same-morning Spotify.

Overnight in AI and cyber for CIOs and CISOs.OpenAI says a ten-thousand-agent swarm on an unreleased model cleared a Clay Millennium math problem — Lean-checked — and there's a scrap over credit. Meta shipped Muse, a personal agent that can reach email and payments, with internal testers already finding photo leaks. Google's threat team tracked criminals using AI agents to steal thousands of credentials in under six hours, including stealers dropped into .claude and .cursor folders. Check Point showed how a shared package service turned ChatGPT into a coerced insider on Gmail — now patched, still an architecture lesson. Cognition hit a forty-eight-billion mark and Mistral took three billion euros from Samsung. Plus an Australian deepfake "slopaganda" network Meta was paying.Hosts: Alice and Tim.No LinkedIn. Same-morning Spotify.

Unsupervised Overnight — Tuesday 8 Sep 2026. Alice and Tim.Anthropic's IPO clock is real enough for the board pack: prospectus slipped to late September, roadshow mid-October, listing aimed days before the US midterms — and they walked away from a six-billion Decart deal. Then the honest Astra vs Fable peer map (plus Nvidia's boss declaring AGI). OpenAI's new chapter on the German wiki: when they tell you. The UN names four labs. Malaysia's central bank fights deepfake governors again.No LinkedIn. CIO/CISO commute briefing.

Monday commute. Four labs shipped in one week and buyers are calling it model fatigue — freeze three private eval tasks, don't chase Thursday's model. OpenAI says coding agents now put about three agent-workdays behind every human researcher day and claims an automated research intern, with March 2028 as the researcher target; ask what they pause when monitoring fails. Astra's launch numbers moved after publish, and the Critical model shows less of its thinking — never take a chart without the harness footnote. Meta's AIRA3 took Kaggle gold and still isn't for sale. Britain is drowning in AI paperwork. Claude Code can finish CI without a click — deny beats hang. Unsupervised Overnight with Alice and Tim.