Tech leaders have long claimed that AI could speed up discoveries in biology and last week Anthropic announced one such result. The finding came from a research group formed by Anthropic earlier this year and took about 21.5 hours to identify. Gene-editing expert Fyodor Urnov of the University of California Berkeley praised the effort but noted that more work is needed.
Anthropic says Claude helped identify a CRISPR-like enzyme system while Roche expands AI use in drug development and Meta faces questions over the privacy of its new Muse agent.
Not all scientists agree on the significance of this discovery which was made using Claude according to Wired. Claude searched through large genomic databases for examples of reverse transcriptases — proteins that copy RNA into DNA. The AI initially found more than 200,000 possible reverse transcriptases and narrowed the list to several thousand that looked new.
Eventually it identified an unusual family of reverse transcriptases containing repeated DNA sequences similar to CRISPR systems. Anthropic is calling this system ART — short for array-associated reverse transcriptases. The AI agent that made the finding also pointed out that it could be a retron — another type of bacterial immune system.
CRISPR and retrons are both bacterial defense systems but CRISPR is more versatile in gene editing than retrons. Seth Shipman of the Gladstone Institutes said the method of discovery was more interesting than what was found. The reverse transcriptase itself had also been identified by researchers in a 2021 paper while Claude appears to have newly connected it with the repeated DNA sequences.
Shipman added that retrons have been used in gene-editing systems so the discovery could still be valuable.
Meanwhile Roche announced new developments in AI-driven drug research at its Pharma Day 2026 event. The company said artificial intelligence is already helping shape decisions in its drug development pipeline. About 40% of pipeline decisions made between late 2025 and mid-2026 included AI or computational input.
Roche also reported that its Phase 3 trial success rate rose to more than 80% in 2026 year-to-date from 65% the year before. The company aims to bring up to 20 new molecular entities to market by 2030 as part of its long-term strategy. Meta introduced its Muse AI assistant this month with a cute mascot named Jolly.
The character has drawn attention for its public appeal but also raised privacy questions. Security researchers identified vulnerabilities involving Muse and Meta has issued fixes and additional safety measures. A Surfshark analysis found that Muse disclosed potential collection of 31 of the 35 data categories listed in Apple’s App Store privacy system.
Muse can connect to services such as email and calendars and can work across personal accounts when users grant access. Meta says Muse does not directly see stored passwords or payment information because credentials are kept in secure storage. By default interactions with Muse may be used to train Meta’s AI models unless users opt out.
One widely reported incident involved tech reviewer Matt Robb whose home address was shared with a buyer on Facebook Marketplace. Robb had granted Muse broad permission to handle the listing but said he did not expect it to share his address without another confirmation. A buyer later arrived at his home.
AI tools are marketed for convenience but they can come with a trade-off in privacy and control. Users may give AI agents access to information that would normally remain spread across different services.
Multiple incidents have sparked debate about the risks of agentic AI and whether stronger security protections are needed before these systems receive broad access to user data. Meta has faced privacy controversies in the past. Its smart glasses have also faced renewed scrutiny over recording practices and allegations involving biometric data.
Another issue involved Muse accessing Apple Messages data even though a tester said he had not granted the expected permission. Tech columnist Jason Aten reported that Muse synced about 187,000 lines from his Messages database. Meta has said Muse can make mistakes or take unexpected actions even though the company says users control which services it can access.
CNET also reported that internal Meta documents showed an earlier plan to use human contractors to help make some appointments or reservations. That feature appears to have been pulled back before broad release. Surfshark found that Muse disclosed potential collection of sensitive categories including location and health information.
Muse also reserves the right under its privacy disclosures to use some collected audio data for purposes such as analytics or personalization.
A Meta representative did not respond to CNET’s request for comment on those privacy concerns. Meta separately says Muse runs inside a dedicated secure virtual machine and that users can change permissions or opt out of having their interactions used for model training.
IMAGE: Source Own work Author Делфина CC4