The outcome they will not publish
A company says viruses its software designed were given to a patient with an antibiotic-resistant infection, and will not say how that patient fared, citing privacy. The record is what that patient was owed.

Listen to this special topic
read by an AI voiceSomewhere there is a person with an antibiotic-resistant infection who was given viruses designed by a computer. On Oct. 20 the company behind them said so. Radical Numerics called the case the world's first clinical use of AI-generated phages. It says it cannot disclose the patient's course or outcome, citing privacy. Its claim rests on a technical report on its own website, not peer reviewed and in no journal. It would not name the patient or the doctors, the New York Times reported.
So a company published a claim about its own patient and withheld the outcome. That is not evidence. It is an advertisement. The federal trials registry held no study sponsored by the company on Oct. 22. A single-patient expanded access treatment need not be registered. But the company describes a compassionate-use pathway while the Times account refers to a trial, and a trial would ordinarily leave a registry record. Neither account resolves which it was. The company's own post concedes that one patient cannot establish general safety or efficacy. Which is why the record matters more than the claim. The question was never whether software will design medicines. It already does. It is whether the first patients are treated where the result gets written down and published.
The published record is thin. A randomized trial in Nature Medicine in June 2025 gave a drug generated by software for scarring lung disease to fewer than 20 patients in each of four groups, for 12 weeks. Side effects were about as common as on placebo. On a secondary measure the top dose gained roughly a tenth of a liter of lung volume where placebo lost a little. Its authors include employees of the company that designed it.
Failure belongs in the record too, and some of it is there. Recursion halted four programs in May 2025, saying of one that "we saw limited tumor shrinkage and clinical activity." Absci dosed the first volunteers the same month in an early trial of an antibody it says its software designed. It expects to enroll about 40 healthy adults. Phage therapy's own evidence is early and mixed. A Belgian-led consortium reported its first 100 cases and about three in four infections improved, though the authors call the data uncontrolled and clearance was far less likely without an antibiotic alongside. A randomized trial in urinary infection after prostate surgery found phages no better than placebo irrigation, though no worse than antibiotics.
Better evidence is coming, slowly. A placebo-controlled trial of inhaled phages in ventilator pneumonia, 184 patients in France, was registered to begin on Nov. 3 and to finish its main measurement at the end of 2029. A randomized, placebo-controlled trial of another phage product has read out, and its sponsor said in January that the Food and Drug Administration agreed the result supports a larger trial, expected in the second half of 2026.
Three risks sit under the headlines and they are not the same risk. The first is ordinary. Unproven products reach patients through pathways that ask for little. The second is accidental, and it is about laboratories and manufacturing rather than models. The National Institutes of Health's Office of Science Policy calls biosafety a shared responsibility. The third is deliberate misuse. Simon Clarke, an associate professor at the University of Reading, said in August that work on designing phages "raises some serious regulatory and safety concerns, to say the very least." Filippa Lentzos of King's College London and three colleagues wrote in the Bulletin of the Atomic Scientists that "the anticipated risk is hypothetical." They write that this "is not a reason to ignore AI or be sanguine about the risks it poses."
The safeguards already exist. For a single patient the Food and Drug Administration requires the treating physician to sign and answer for the request, a review board's approval before treatment, and informed consent. On misuse, the federal artificial intelligence plan tells federally funded institutions to buy from suppliers that screen orders and verify customers, as the Johns Hopkins Center for Health Security reads it. Anthropic publishes what it calls a voluntary framework for catastrophic risks. It says it can no longer make a strong argument that a given model's biological risks are low, and commits to publish risk reports every three to six months, with outside review in certain circumstances.
Doctors who treat these infections will say a dying patient cannot wait for a trial that finishes measuring in 2029. They are right. Compassionate use exists for that patient and should. Company officials told the Times that the Food and Drug Administration had authorized the work, and the newspaper did not independently confirm it. If an authorization exists, the pathway did its job. No rule in it requires that the result be published. That gap closes with a condition on the case, not a new statute.
So the position is safety first, and safety first turns out to be clerical. A patient who asks should be told the truth. A company says software-designed viruses have reached at least one person. Nothing has been published about how that person fared. No trial has yet tested whether software-designed phages work. A medical staff asked to approve such a case should hold the agency's written authorization for this patient and product, and the review board's approval, not a colleague's nod. Someone should be able to name who made it, and to what standard. The plan should say what will be measured and who records it. And the sponsor should agree in writing, before the first dose, to publish the case and its outcome without naming the patient.
One patient proves almost nothing. The record is what that patient was owed, not a favor to researchers. Medicine's first rule is to do no harm, and it reaches the tools as surely as the hands holding them. A company that claims a first patient and keeps the outcome has done the easy part and skipped the rest.
- Radical Numerics' own account, Oct. 20, 2026In a post dated Oct. 20, 2026, Radical Numerics said that designs generated by its model "were administered to a patient" and called this "the world's first clinical use of AI-generated phages." The post says the patient had "an antibiotic-resistant bacterial infection" and that the treatment was "authorized, physician-led" and carried out under a compassionate-use pathway, naming no hospital, physician, trial registration, case report or preprint of this patient's treatment. It states: "we cannot disclose the patient's clinical course or outcomes to safeguard their privacy"; "observations from a single patient cannot establish general safety or efficacy"; and that such investigational therapies "have not been approved by regulatory bodies for broader clinical use" and are "subject to oversight and safety requirements established by the relevant government regulatory authorities." On its own practices the company says it is "committed to rigorous precautions and procedures for safe wet-lab research."
- What the public record containsAn account in The New York Times, dated Oct. 20, 2026 and republished by GV Wire, reports that the company's claim rests on a technical report on its own website that has not been peer reviewed or published in a scientific journal, that company officials said the Food and Drug Administration had authorized the work, and that the company declined to share information about the patient or the treating doctors, citing confidentiality. The account does not independently confirm the authorization. It refers to a trial and to the doctors running it, while the company describes a compassionate-use pathway, and neither account resolves which it was. The account does not say that reporters reached the treating clinicians. A search of ClinicalTrials.gov on Oct. 22, 2026, reproduced on Nov. 8 and Nov. 17, 2026, returned no study with Radical Numerics as sponsor. Four expanded access records for bacteriophage products are registered there (NCT03395769, NCT03395743, NCT06798168 and NCT04636554), sponsored by other organizations. Single-patient expanded access use does not require registration.
- The randomized trial of a drug generated by AINature Medicine published the first phase 2a trial of rentosertib (formerly ISM001-055) on June 3, 2025: a multicenter, double-blind, randomized, placebo-controlled study in idiopathic pulmonary fibrosis. Patients received 12 weeks of 30 mg once daily (n=18), 30 mg twice daily (n=18), 60 mg once daily (n=18) or placebo (n=17). The primary endpoint, the share of patients with at least one treatment-emergent adverse event, was 72.2%, 83.3%, 83.3% and 70.6% respectively. Among the secondary endpoints, mean change in forced vital capacity was +98.4 ml (95% confidence interval 10.9 to 185.9) at 60 mg once daily against -20.3 ml (95% confidence interval -116.1 to 75.6) on placebo. The commonest events leading to discontinuation related to liver toxicity or diarrhea. Employees of Insilico Medicine, which discovered the drug, are among the authors. Registration NCT05938920.
- AI-discovered molecules in the clinic, and the ones that stoppedAn analysis of AI-native biotechnology pipelines published in Drug Discovery Today on April 30, 2024 by five Boston Consulting Group authors reports: "In Phase I we find AI-discovered molecules have an 80-90% success rate, substantially higher than historic industry averages" and "In Phase II the success rate is ~40%, albeit on a limited sample size, comparable to historic industry averages," that is, about two in five. Absci announced on May 13, 2025 that "the first healthy volunteers have been dosed" in a phase 1 study of ABS-101, an anti-TL1A antibody for inflammatory bowel disease the company says was "designed and optimized using our generative AI platform," adding that "The study is expected to enroll approximately 40 healthy adult participants." Genetic Engineering and Biotechnology News reported on May 5, 2025 that Recursion ended four programs and paused a fifth, quoting its chief executive, Chris Gibson, on REC-994 in phase 2, "without a definitive signal, we did not think it prudent to go forward," and on REC-2282 in phase 2, "we saw limited tumor shrinkage and clinical activity."
- Where phage therapy's own evidence standsNature Microbiology published on June 4, 2024 a retrospective observational analysis of the first 100 consecutive cases of personalized phage therapy arranged by a Belgian consortium in 35 hospitals in 12 countries between 2008 and 2022. Clinical improvement was reported for 77.2% of infections and eradication of the targeted bacteria for 61.3%. Eradication was 70% less probable when no concomitant antibiotics were used (odds ratio 0.3; 95% confidence interval 0.127 to 0.749). Fifteen adverse events were reported, including seven non-serious suspected adverse drug reactions. The authors write that the analysis "is limited by the uncontrolled nature of these data." In a randomized, placebo-controlled, double-blind trial published in The Lancet Infectious Diseases, 113 men scheduled for transurethral resection of the prostate were randomized, and 97 received at least one dose and were included in the primary analysis; urine culture normalized in 5 of 28 on phages (18%), 9 of 32 on placebo irrigation (28%) and 13 of 37 on antibiotics (35%). The trial concluded that phage therapy "was non-inferior to standard-of-care antibiotic treatment, but was not superior to placebo bladder irrigation."
- The trials that could settle itPYOPHANEB (NCT07804082), sponsored by Assistance Publique, Hopitaux de Paris, is a phase 3 double-blind, placebo-controlled study of nebulized bacteriophages added to intravenous antibiotics in ventilator-associated pneumonia, with 184 planned participants at hospitals in France. The registry record lists a start date of Nov. 3, 2026 and primary completion Dec. 3, 2029, with status not yet recruiting. Armata Pharmaceuticals said on Jan. 13, 2026 that the Food and Drug Administration agreed data from its completed phase 2a "randomized, double-blind, placebo-controlled" study support advancement to a phase 3 study, which "is anticipated to initiate in the second half of 2026." Separately, Locus Biosciences announced on Jan. 7, 2026 a contract award from the National Institute of Allergy and Infectious Diseases of $3.3 million, with up to $28 million if all milestones are met, under contract No. 75N93025C00035, to fund a phase 1b trial of what it describes as "an AI-designed bacteriophage therapeutic for hospital-acquired pneumonia."
- The oversight that already appliesThe Food and Drug Administration's guidance on single-patient expanded access states that "you, the physician, must ultimately sign and be responsible for the request"; that "you must obtain approval from an Institutional Review Board prior to treating a patient under expanded access," with concurrence of the board's chair or a designated member available as an alternative on Form FDA 3926; that informed consent must be obtained from the patient or a legally authorized representative; and that the agency has "30 calendar days to review the request, but in most cases, we complete the review within a few days." For emergencies: "You must report the emergency treatment to an Institutional Review Board within 5 working days" and "submit the written request to FDA within 15 working days of the initial authorization." The National Institutes of Health Office of Science Policy, in a post last updated Feb. 28, 2025, states that "biosafety is a shared responsibility for all those involved in the research enterprise."
- Screening policy and the developers' own commitmentsThe Johns Hopkins Center for Health Security's guide to the federal AI Action Plan, written by Melissa Hopkins, reports that the plan directs institutions receiving federal research funding to use nucleic acid synthesis tools and "synthesis providers that have robust nucleic acid sequence screening and customer verification procedures." The Institute for Protein Design at the University of Washington says more than 170 scientists who lead research teams signed community standards in 2024 "creating obligations to report concerning research practices" and committing them to "purchase synthetic DNA only from providers that adhere to industry-standard biosecurity screening practices." Anthropic's Responsible Scaling Policy, version 3.0, published Feb. 24, 2026, describes "the voluntary framework we use to mitigate catastrophic risks from AI systems" and says of biological risk that "we can no longer make a strong argument that risks are low from a given model." It requires external review in certain circumstances and publishes risk reports every three to six months. It is one developer's policy, not an industry rule.