Anthropic Says Claude Found a CRISPR-Like Phage Enzyme System With Unknown Function

01Anthropic Says Claude Autonomously Discovered a CRISPR-Like Enzyme System, but Its Function Remains Unconfirmed

Anthropic says Claude has identified a previously uncharacterized enzyme system in bacteriophages, the viruses that infect bacteria, after searching a large database of DNA sequences with only high-level direction from company scientists. The finding came from one of the first programs at Anthropic’s new life sciences research group and laboratory, formed in spring 2026 to test whether AI agents can participate in biological research from data analysis through experimental validation.

The search began with a broad prompt: find interesting new examples of reverse transcriptases, enzymes that copy RNA into DNA. Roughly 950 Claude agents then spent 21 hours and 210 million tokens examining the database, investigating distinct reverse-transcriptase families and choosing which candidates appeared worth pursuing.

According to Anthropic, one agent noticed an unusual repeating pattern of DNA beside the gene for an odd-looking reverse transcriptase. The underlying enzyme, found in a jumbo phage, had already appeared in earlier research. Anthropic says Claude’s contribution was connecting it to an array of non-coding DNA sequences and to an additional protein whose function is unknown.

Human researchers handled the subsequent analysis and all laboratory work. After testing the candidate, the team classified it as a previously uncharacterized phage enzyme system and named the family “array-associated reverse transcriptases,” or ARTs. Anthropic has published the results in a preprint, meaning the account of Claude’s autonomy and the system’s novelty comes from the company and its researchers; the source provides no independent replication.

ART drew comparisons with CRISPR because it is associated with an array of repeated DNA sequences. CRISPR itself was first recognized through unusual repeats in bacterial DNA before becoming the basis of gene-editing medicines. Anthropic also says ART combines characteristics previously found together in only a handful of other systems, all of which are programmable and perform operations such as cutting, copying or pasting DNA.

That resemblance does not establish ART as a gene-editing tool. Anthropic has not determined the system’s primary biological function, and the reported experiments do not show that it can cut, copy or paste DNA in the manner of known programmable systems. The company says work to understand ART is continuing.

02Microsoft-commissioned survey across 37 countries finds 68% of daily US AI users still worried

Daily use of artificial intelligence does not necessarily bring confidence in the technology. A new Gallup survey commissioned by Microsoft found that 68% of Americans who use AI every day are worried about it, with concern even higher among people who use it less frequently.

Gallup surveyed about 1,000 people in each of 37 countries between April and July as part of a project eventually intended to cover 140 countries. The findings arrive amid debate over AI safety and the technology’s effects on jobs and society, complicating the idea that adoption alone measures public enthusiasm.

The US results combine high use with high concern and limited trust. Among all American respondents—not only daily users—74% expressed worry, while just 36% expected AI to mostly help the country. Among daily US users, 45% said they trusted AI’s results “a lot.”

Those figures measure different things and have different denominators. The 68% figure covers worry among daily American users; the 74% figure covers worry among all US respondents; and the 45% figure measures daily users’ confidence in AI outputs, not their view of its wider social effects. Gallup senior scientist Pablo Diego-Rosell said people can use AI frequently, expect benefits and remain worried while also doubting whether its information is accurate.

The contrast with Singapore was pronounced. Singapore had the survey’s highest share of daily AI users, at 46%. More than 80% of respondents there who knew about AI expected it to improve everyday life, while 77% thought it would help the country. Across all 37 countries, the median share of respondents who strongly trusted AI results was 36%.

The broader sample was less anxious than the US. Curiosity was the leading sentiment in 34 of the 37 countries, and about 32% of respondents across the surveyed countries reported worry.

The results should not be treated as globally representative or as proof that AI use causes either concern or optimism. India, Australia and Malaysia were absent from the published results, and attitudes may have changed since fieldwork ended in July.

03Linear Reworked CI After Tests Nearly Quadrupled, Cutting PR Waits to About 5 Minutes and Halving Per-Test Runner Use

Linear has reworked its continuous integration system after AI coding agents accelerated code production faster than the company could validate it. Continuous integration, or CI, automatically checks proposed code changes; at Linear, every pull request must pass those checks before it can be merged.

The company said its test suites have almost quadrupled since the start of the year, increasing infrastructure use and leaving developers and agents waiting for feedback. Linear optimized for two measures: the time a pull request waits on CI and the runner time consumed by each test. It reduced the former from more than six minutes to just over five minutes and cut the latter roughly in half.

Linear first moved its workloads from GitHub Actions to unnamed third-party runners with faster processors, storage and caching. In a like-for-like comparison covering the two days on either side of the switch, jobs ran 34% faster on average. Separately, adopting tsgo, the native TypeScript compiler, reduced the weekly median duration of its TypeScript check by 73%.

Linting had also been rebuilding the complete TypeScript type graph because several custom rules depended on type information. Linear rewrote those rules to analyze syntax instead, cutting API lint time by 68% and repository-wide lint time by 55%, while substantially reducing memory use.

The team then targeted small jobs that blocked all later work. Limiting how much repository history change-detection jobs fetched reduced the slowest gate from 94 seconds to 20 seconds. Removing checkout from jobs that did not need a working tree cut their time from 27 seconds to seven seconds. For events that still required path comparisons, sparse, blobless checkouts with limited history saved about another 11 seconds.

Linear also replaced its checkout action with one that retries failed transfers, aborts stalled connections after about 30 seconds and uses a persistent repository cache. Moving cache-marker writes out of the merge-critical path removed 42 seconds from every API pull request and merge-queue entry. Together, the critical-path changes saved roughly one minute on API pull requests with cache misses and reduced runner starts.

The broader techniques—faster runners, modernized compilers, syntax-only lint rules, smaller checkouts and removal of nonessential gates—can transfer across technology stacks, but Linear did not disclose total savings or establish that its internal results will reproduce elsewhere.