"The AI Pulse — 2026-10-31: A Designed Virus In, Eleven Billion Out"

The same 72 hours produced two firsts pointing in opposite directions: a machine-designed virus went inside a human patient for the first time, and the market that funds such machines said no to eleven-nineteenths of a $30 billion ask. One story is biology learning to compile; the other is capital learning to count. The history of AI's rise will be written in both registers, and today they arrived together.

The first AI-designed therapy went into a patient

Radical Numerics' Omnii-1 genome model designed bacteriophages — viruses that hunt bacteria — which killed a patient's antibiotic-resistant isolate in vitro and were then administered under compassionate use. Before this, phage therapy meant digging through sewage and soil libraries for a natural virus that happened to match your infection, a matching process measured in months. The Omnii-1 result flips the frame: instead of finding the virus, design one, with a genome language model predicting which designs will actually infect the target strain. The caveats matter and are real — outcomes undisclosed, no peer review yet, a company announcing its own milestone — but the category crossing is not a caveat: this is the first AI-designed biological entity dosed into a human, and the barrier between "AI assists drug discovery" and "AI authors the therapeutic" just stopped existing. radicalnumerics.ai

OpenAI raised $11 billion of a $30 billion ask

The WSJ's sources say the round that was pitched at $30B/$1.4T closed this morning at roughly $11 billion in commitments and a $1.25 trillion valuation — less than half the target on both axes. The tell is in the pivot: OpenAI reportedly now plans to chase family offices and high-net-worth individuals for the balance. When institutions with research teams pass and you go hunting for retail-adjacent money, the "capital is infinite" assumption underpinning the entire scaling race has its first hard datapoint against it. Note what did not break: $1.25T is still the largest private valuation ever. The market isn't refusing to fund the frontier; it has started price-discriminating between the frontier and the finance plan attached to it, which is a very different — and much earlier-stage — signal than a bubble popping. ttm.financial (Dow Jones/WSJ text)

Blackstone prices Anthropic's debt as junk

Per an FT report, Blackstone has told investors Anthropic will land a high "junk" rating — the top tier of speculative grade — when it goes public, not investment grade. The delicious detail is who's saying it: Blackstone is the anchor investor in the $63B chip-financing package, so the same institution funding the compute buildout is simultaneously marking the debt that funds it as speculative. That's not contradiction; that's what mature markets do with risks they take seriously — they buy the equity and short the credit. The $80.7B H1 loss already in Anthropic's IPO filing makes the agencies' math legible: this is a bond-market verdict that AI labs are growth companies with telecom-grade capital intensity, a combination the credit side of the market has seen end badly before. techcrunch.com

Northflank makes nine: the AI dev stack became a breach genre

Northflank's own status page told customers to assume every environment variable — every API key, secret, and credential — was decrypted and gone, and to rotate everything. Wiz's incident tally now counts nine AI developer platforms disclosing breaches in October alone. The reason this is a genre rather than bad luck: the AI tooling stack concentrates exactly what attackers want — long-lived credentials for frontier-model APIs and cloud accounts, often with billing attached — in platforms whose entire pitch is automating away the friction that used to slow humans down. Nine platforms in a month means the comfortable assumption "my secrets live safely with my provider" is now operationally dead, and credential rotation is about to become a recurring line item in every AI engineering budget the way backups did after the last era's ransomware lesson. status.northflank.com — analysis: qovery.com

GPT-6 Rosalind loosens the bio boundary behind an access list

Buried in a developer changelog — the quietest possible venue for the loudest possible policy question — OpenAI shipped GPT-6 Rosalind, a life-sciences reasoning model on GPT-6 Astra weights with a deliberately looser dual-use biology boundary than its parent, gated only by a trusted-access program. One week after the UK opened its AI Biosecurity Lab specifically to study this exact seam, the policy experiment biosecurity people have been war-gaming for two years is now live, and the bet inside it is legible: OpenAI is wagering that screening who can use a capable model substitutes for limiting what the model helps with. If the access list holds, expect every lab to copy the pattern; if it doesn't, the incident will define the other template. Either way, the decision arrived as a changelog entry rather than a policy debate, which tells you where these calls are actually being made. developers.openai.com — context: datastudios.org

Watch-item for tomorrow: Alibaba's Cloud CTO has "Qwen4 next week" on stage at Apsara Amsterdam, with four variants teased (llms.blog) — days after Mistral Large 4's open weights took the open-model crown, the open-weights race is about to get a second flagship inside a week. Two days after a $19B funding gap, someone should ask what a training run costs when your competitors give the weights away.