Artificial intelligence
Anthropic's Agents Found a CRISPR-Like System in 21 Hours. Its CEO Took the Result to the UN Security Council
Anthropic says roughly 950 Claude agents surfaced a previously undescribed enzyme system in phage genomes, and Dario Amodei cited it the same day while asking the UN Security Council to ban AI's use in bioweapons development. The enzyme's function is unknown, and one report says the search could not be reproduced.
MAI
Anthropic published a preprint on 23 September claiming that Claude agents, running largely unattended, identified a previously undescribed enzyme system buried in bacteriophage genomes. The company calls it ART — array-associated reverse transcriptase. The same day, chief executive Dario Amodei sat before the UN Security Council at its first meeting devoted to the safety risks of advanced AI and asked the Council to start with a narrow international agreement: a ban on using AI to develop biological weapons. He cited the enzyme result by name, and called it preliminary.
The two events are not in tension. Read together, they are the clearest statement yet of what Anthropic thinks its own technology is for and what it thinks the technology is capable of doing in the wrong hands.
What the agents actually did
The search was industrial rather than clever. Anthropic set roughly 950 Claude agents loose on a large database of genomic sequence, gave them tooling, and let them run.
| Measure | Reported figure |
|---|---|
| Claude agents | ~950 |
| Wall-clock search time | ~21 hours |
| Tokens consumed | ~210 million |
| Reverse transcriptases collected | 200,000+ |
| Candidate systems flagged | ~3,500 |
| Candidates taken to close analysis | 20 |
| Novel systems described in the preprint | 1 |
The system that survived has three parts: a reverse transcriptase, an adjacent partner gene, and an array of evenly spaced DNA repeats — between three and twenty-one copies, depending on the example. That architecture is what makes the comparison to CRISPR obvious. The moment of recognition, as the preprint records it, came from an agent reading sequence directly: "The DNA next to the RT is spectacular: I can see by eye a tandem repeat array…that's a CRISPR-like…repeat array?!"
Early bench work by human scientists shows the array is transcribed into distinct short RNAs, which appear in quantity soon after a phage infects its host. That is the extent of what is established.
The word doing the work is "like"
What ART does biologically is unknown. Anthropic says so plainly, and says experiments are ongoing. "CRISPR-like" is a claim about shape — repeats, spacing, an associated enzyme — not about function. CRISPR became consequential because the repeats turned out to encode targeting information that a nuclease could be pointed at anything. Nobody has shown that ART's repeats do anything comparable. A system that merely looks like a programmable one is a research lead, not a tool.
Feng Zhang of MIT and the Broad Institute, who has better reason than most to be sceptical of CRISPR comparisons, gave the finding a careful endorsement:
This is an exciting example of how AI agents can contribute to biological discovery. The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation.
That is praise for the method, not a verdict on the molecule.
There is a further wrinkle that deserves more attention than it has received. The Next Web, reading the preprint, reports that ten subsequent identical searches failed to surface the same array again, a result Anthropic attributes to the scale of the search space and the unpredictability of agents. If that holds, the pipeline is not yet a repeatable instrument. It is a lottery with good odds and a very large number of tickets, and the scientific value of a result you cannot re-derive on demand is limited until someone can.
The capability is the search, not the enzyme
Strip out the CRISPR framing and what Anthropic has demonstrated is a throughput claim: a week or more of a specialist's comparative genomics, compressed into a day, for the price of 210 million tokens. That is genuinely useful, and it is also the uncomfortable part.
A system that can sweep two hundred thousand enzymes and rank the interesting ones does not care what makes them interesting. The same pipeline pointed at a different objective is exactly the capability Amodei described to the Security Council when he asked for a bioweapons agreement with verification attached — nations able to check each other's compliance, common standards for testing models, and a duty to report incidents. Anthropic notes that all wet-lab work on ART was done by human scientists at biosafety levels 1 and 2, with no human pathogens involved. That constraint is a policy choice, not a property of the software.
Announcing the discovery on the day of the UN appearance is therefore not a coincidence but an argument: here is what this does for medicine, and here is why you should write the rules now. Whether the Council treats a preliminary, unreviewed, not-yet-reproducible finding as evidence for either half of that case is a separate question.
Sources: Anthropic: Claude discovers a novel enzyme system with CRISPR-like repeats · The Next Web: Anthropic says Claude found a new enzyme system with CRISPR-like repeats · The Next Web: Dario Amodei asks UN Security Council to back a ban on AI bioweapons · Unite.AI: Anthropic says Claude discovered a new enzyme system resembling CRISPR · Malay Mail / AFP: Anthropic says its Claude AI helped uncover previously unknown enzyme system · Interesting Engineering: Claude scans 200,000 enzymes to uncover new CRISPR-like system hidden in phages