Anthropic’s bio cases, Mythos for Europe, and OpenAI’s Pro pause

In this Brief

• Anthropic Sep TI: biological misuse cases + China distillation ~200M
• ENISA got Mythos 5 late, still not Mythos 5.1
• Hawley Senate probe of OpenAI Hugging Face breach
• OpenAI pauses new $200 Pro / Astra capacity
• Oracle cloud infra +121%; Amazon×OpenAI ChatGPT ads pilot

Anthropic's bio cases and the distillation ledger

Setup: Anthropic's September 2026 threat-intelligence report covers operations disrupted from Dec 2025 to Aug 2026 across seven harm areas, including biological misuse and illicit distillation. Models named: Claude Haiku, Sonnet, Opus. None of the misuse cases involved Fable or Mythos-class models except one illicit distillation case.

Ian Carlos Campbell at Engadget writes that the report includes five biological-misuse case studies. Anthropic told the NYT (via Engadget) that vaccine-style work can look like bioweapon work; the company erred toward caution and shut accounts without asserting proven intent. Jacob Klein, Anthropic's head of threat intelligence, via Engadget: it is "an incredibly nuanced situation," not a comic-book villain asking for a weapon.

Russell Brandom at TechCrunch covers the same report's distillation chapter: nearly 200 million exchanges across five campaigns. Distillation here means harvesting a frontier model's answers and hidden step-by-step reasoning to train a smaller student model. Alibaba: 151 million exchanges May to July 2026, peak nearly 3 million/day across about 3,500 accounts, framed as training material for Qwen. Moonshot (Kimi): nearly 300,000 requests over 10 days via about 5,000 accounts, primarily targeting Opus.

Take: The biology cases are dual-use scientists, work that can help or harm depending on who uses it. The industrial story is rivals vacuuming Claude's reasoning at warehouse scale. Both are Anthropic's own case file, disclosure with the company's framing still attached.

Longer cut: Anthropic flagged scientists using Claude on dual-use biological research and shut the accounts. Live X: live Aaron_Harme X (Thu evening; status URL not on disk).

ENISA got Mythos 5, still not 5.1

Setup: Euronews reports that Anthropic granted the EU's cybersecurity agency ENISA access to Mythos 5 for testing now, more than three months after the model was already in circulation for others. Mythos 5.1 remains withheld from that channel. The UK AI Security Institute was separately left out of Mythos 5.1 pre-release access earlier in the week.

Access geography is the story: which sovereign tester gets which weight class, and how late.

Take: Europe got yesterday's frontier under glass. Britain is still arguing about the newer one. Pre-release access is the real safety policy.

Longer cut: ENISA got Anthropic's Mythos 5 more than three months late, and still not Mythos 5.1. Live X: live Aaron_Harme X (Thu noon; status URL not on disk).

Hawley's Hugging Face letterhead

Setup: Axios reports that Sen. Josh Hawley's Senate Disaster Management subcommittee opened a probe of OpenAI's Hugging Face breach, with an Oct 1 deadline and 16 questions on the letter. Hugging Face is a major public hub for hosting and sharing AI models; the breach story is the agent-escape / security-control failure OpenAI has been answering for since August.

Take: Putting a model-security failure on a Disaster Management letterhead is the tell. Congress is reaching for incident language, not just product language.

Longer cut: Hawley put OpenAI's Hugging Face breach on a Disaster Management letterhead. Live X: live Aaron_Harme X (Thu morning; status URL not on disk).

OpenAI pauses new $200 Pro sign-ups

Setup: Sarah Perez at TechCrunch writes that demand for Astra, OpenAI's newest high-capability model line launched Sep 3, led the company to temporarily pause new subscriptions to the $200/month Pro plan. Product leader Thibault (Tibo) Sottiaux: Pro puts the most strain; sign-ups for that tier disabled so existing users keep usable access.

API and lower-cost Go / Plus plans remain available. Existing Pro accounts unaffected. Duration and daily signup volume not disclosed.

Take: This is a capacity bill arriving in public on the most expensive consumer tier, not a safety pause.

Live X: live Aaron_Harme X (Thu evening; status URL not on disk).

Oracle's cloud bill and ChatGPT ads

Setup: Jordan Novet at CNBC reports Oracle Q1 revenue $19.35B, cloud infrastructure $7.4B (+121%), with the stock jumping after hours. The same print carries the cost side of the AI landlord boom: heavy capital spend and deeply negative free cash flow in the coverage cluster.

Marketing Dive reports Amazon is piloting ad services inside ChatGPT via Amazon's ad platform (DSP) in the US, with ChatGPT Ads pitched toward $1B in annual recurring revenue and Amazon's ads business around $70B the prior year.

Take: The picks-and-shovels quarter and the chatbot ad pilot are the same economy from two angles: who rents the racks, and who rents the attention.

Live X: live Aaron_Harme X (Thu evening Oracle; Thu noon Amazon ads; status URLs not on disk).

Also noted

TechCrunch writes that Paul Christiano joined the OpenAI Foundation board and Safety/Security Committee and says the industry, including OpenAI, is not on track to reduce catastrophic loss-of-control risk. Live X: live Aaron_Harme X (Thu noon).

PR Newswire says Positron raised $875M Series C at a $5B valuation after the round for memory-first inference silicon (Asimov/Titan; LPDDR5X; 50+ Atlas racks at OCI). Live X: live Aaron_Harme X (Thu noon).

Hugging Face lists DeepSeek-V4.1-Flash as a 552B mixture-of-experts model with 1M context. Live X: live Aaron_Harme X (Thu morning).

The Next Web reports a DOJ inquiry into whether Nvidia's Groq licence-and-hire structure avoided antitrust review. Live X: live Aaron_Harme X (Thu morning).

The Wednesday Brief covered Anthropic's four Claude cyber breakouts + METR, Britain's Mythos 5.1 snub, Coxon/Hubinger, Apple Watch Audio Intelligence, and Tailwind×Shopify.

Sources

• Anthropic: September 2026 threat intelligence report
• Engadget: biological misuse cases
• TechCrunch: China distillation campaigns; OpenAI Pro/Astra pause; Christiano board
• Euronews: ENISA Mythos 5 access
• Axios: Hawley Hugging Face Senate probe
• CNBC: Oracle Q1
• Marketing Dive: Amazon×OpenAI ChatGPT ads pilot
• PR Newswire: Positron $875M
• Hugging Face: DeepSeek-V4.1-Flash card
• The Next Web: DOJ Nvidia×Groq