OpenAI backed a startup built to fight an AI-made pandemic
A biosecurity startup raised $36 million to design antibodies against the pathogens AI might help build — and it was already quietly firing at the attack surface it is defending. Separately, DHS started using AI to recommend what to redact from the records you asked for, and an anonymous model took third place on the world's busiest token leaderboard with a fingerprint no one was supposed to see.
Red Queen Bio, a biosecurity startup whose $15 million seed OpenAI led in November, has now raised $36 million in total, per the Wall Street Journal, and is working on antibody countermeasures against pathogens a future AI could help engineer. The founder pedigree is not incidental: founders Hannu Rajaniemi and Nikolai Eroshenko spun Red Queen out of HelixNano, and the origin story is openly defensive — the same models that are getting better at designing proteins are getting better at a dual-use problem, so the bet is that you build the antibodies before you need them. We covered the evidence for that reading earlier this month — GPT-6 Astra outscores every specialist at antibody prediction. The timeline is the soft part: antibody countermeasure work usually takes years, and the human-trial target for influenza-family threats — next year — is the company's intention, not a schedule.
The part that gives the claim teeth was published by OpenAI itself: an evaluation where Red Queen ran a wet-lab experiment in which a GPT-5 model improved molecular cloning efficiency 79 times by discovering a novel recombination mechanism. Read plainly, that is the industry's own demonstration that frontier models are now measurably better laboratory hands than the research staff. Red Queen's entire thesis is what happens when the same capability falls to someone with worse intentions. The premium here is not paid for the platform quality; it is paid for who arrived first.
OUR TAKE: The strategic read is that defence is racing to out-pace the model capability curve rather than predict it. Digitization of pathogen design doesn't have a bottleneck issue — protein design and synthesis have both reached the point where the limiting factor is neither knowledge nor compute. Red Queen's bet is that the response also has to be AI-native, and this funding round is the market paying insurance premium for that. It is also why OpenAI put itself on the cap table at seed: the lab that shipped the capability has a documented interest in the countermeasure existing, and the arrangement keeps the countermeasure where the lab can see it.
The US Department of Homeland Security will use AI to recommend what to redact from FOIA records, rolling out through the request types that make up its heaviest volume. Customs and Border Protection's RedactAI, built on Google's Vertex AI on Google Cloud Platform, is listed in DHS's own AI use-case inventory as pre-deployment and explicitly flagged as not high-impact. DHS also posted a notice of intent to award Deloitte a contract for customized FOIA software. The automated tooling is applied to request types above 10% of FOIA volume, which a DHS spokesperson put at approximately 100,000 to 140,000 requests. Whether DHS over-redacts that massive a slice of the public record is the fight waiting to happen. Transparency groups are already saying so: the University of Florida's Brechner Center called it "a very powerful tool for secrecy," EPIC said it "almost guarantees that DHS will over-redact," and the ACLU's position is that the technology simply is not ready.
An unidentified model is sitting third on OpenRouter's global token leaderboard, and the way the internet found its owner was the tokenizer. "Space Bunny Alpha," published by the anonymous account "stealth," processed 13.9 trillion tokens in the week of September 21 to 27 — third behind DeepSeek V4.1 Flash at 19.6 trillion and Zhipu's GLM 5.3 Flash at 16.3 trillion, according to the weekly tally the Chinese business press publishes from the OpenRouter data. Developers noticed that token counts on 50 test strings matched MiniMax's models exactly, and MiniMax opened a beta for its own M3.1-Flash preview on September 28. MiniMax has not claimed the model; the tokenizer is the entire case.
Separately on the same scoreboard: China's model call volume led worldwide for a 22nd consecutive week at 62.22 trillion tokens, even as the global total rose 13.18%, and last week's no. 6 (Xiaomi's MiMo-V2.5) and no. 7 (Tencent's Hy3) fell off the list.
What to watch: whether MiniMax confirms the Space Bunny Alpha attribution, and whether DHS publishes redaction-accuracy numbers for RedactAI the way it publishes its FOIA backlog.
As agencies hand first-pass redaction decisions to models, would you rather see them publish accuracy numbers before rollout or accept a human in the loop as sufficient oversight? Tell us in the comments.
Sources: Wall Street Journal · Reuters · OpenAI · Red Queen Bio · AbTherx · FedScoop · DHS AI Use Case Inventory · Washington Post · OpenRouter · The Paper · National Business Daily
