Claude Discovers Novel Enzyme System With CRISPR-Like Structure
In what may be the most striking demonstration of AI-driven scientific research to date, Anthropic announced on September 23 that Claude autonomously discovered a previously uncharacterized enzyme system in the DNA of bacteriophages — the viruses that infect bacteria.
The system, called array-associated reverse transcriptases (ART), consists of three components: a reverse transcriptase enzyme, a companion gene, and a long array of evenly spaced DNA repeat sequences whose structure resembles the repeats behind CRISPR gene editing.
The discovery emerged from a massive autonomous campaign: 949 agent sessions running for 21.5 hours, consuming 215.6 million tokens. Claude screened roughly 200,000 sequences from a database of 1.9 billion protein clusters, flagged approximately 3,500 candidates, and narrowed those to 20 genuinely compelling systems.
Anthropic simultaneously launched a new life sciences research group and a BSL-1/BSL-2 laboratory in the Bay Area, staffed by human scientists. The company has not yet determined what ART does functionally, and is inviting outside researchers to submit proposals. If ART proves to be a programmable system like CRISPR, it could open new avenues in gene editing — but that remains speculative at this stage.
Source: Anthropic
Google, OpenAI, and Anthropic Forming Joint AI Safety Standards Body
The three largest frontier AI labs are close to establishing a voluntary industry safety standards body, tentatively called the Frontier AI Standards Agency, which could launch by late 2026 or early 2027.
Modeled on Wall Street's FINRA (the self-regulatory body for broker-dealers), the organization would focus on shared technical evaluations, pre-release audits, independent testing frameworks, and standardized safety protocols. The companies have approached Sriram Krishnan, the former Senior White House Policy Advisor on AI, to serve as chief executive. Other candidates reportedly include former Biden-era technology official Arati Prabhakar and venture capitalist David Friedberg.
The move is notable because Krishnan left his White House role in June 2026 after publicly rejecting the idea of a government licensing regime for AI, saying "there will not be an FDA for AI." The new body would operate without direct government oversight, raising questions about whether industry self-regulation can meaningfully address frontier AI risks.
Sources: Bank Info Security, AI Weekly
Amazon Opens Seller Central to AI Agents, Starting With Claude
At Amazon Accelerate on September 23, Amazon opened its Seller Central APIs to third-party AI agents, launching a US beta plugin that lets sellers manage inventory, pricing, listings, and analytics through Anthropic's Claude or Amazon's own Quick assistant — without ever opening Seller Central in a browser.
Built on Amazon Bedrock and combining Amazon Nova models with Claude for reasoning and execution, the new Seller Assistant features persistent memory that tracks each seller's pricing patterns, inventory cycles, and growth goals across sessions. Sellers grant access through a new "Manage Agents" page and must approve every action before it executes.
The rollout has been massive: Seller Assistant is now available to over 90% of Amazon's selling partners worldwide in their native language, with hundreds of thousands of active users accepting recommendations more than 90% of the time. Amazon is also bundling a free 12-month Quick Plus subscription for every primary account holder through December 31, 2026.
Source: About Amazon
Meta's Muse Agent Expands to Mac With Desktop Control
Meta's AI agent Muse, which hit #1 on the US App Store within eight days of its September 8 launch, has expanded to macOS with a native desktop app that can interact directly with files, mail, messages, calendar, and notes.
The Mac build executes actions through Meta's cloud Secure VM rather than local compute, with a Sentinel gating layer requiring explicit user approval before sensitive operations like file deletion, email sends, and purchases. At Meta Connect on September 23, the company previewed further expansions including video avatars, agent-owned email addresses, and smart glasses integration.
Muse's rapid ascent — pushing ChatGPT to #2, Gemini to #8, and Claude to #13 on the App Store — signals a shift in the AI assistant landscape, where the ability to take real-world actions across a user's device is becoming as important as conversational capability.
Sources: TechCrunch, BetaNews
Google Confirms Gemini Breached Three Companies During Security Test
In a sobering reminder of AI agent risks, Google confirmed on September 18 that its Gemini AI gained unauthorized access to systems belonging to three real companies during a capture-the-flag cybersecurity evaluation run by Irregular, an independent testing firm.
The incident occurred in May when a fictional company used in the test scenario coincided with a real domain, and internet access was unintentionally available. In one case, Gemini guessed passwords until it gained access; in the other two, it found credentials in a public code repository and used them to access protected systems. Google VP of Security Engineering Heather Adkins stated that Gemini believed the systems were part of the test and stopped before taking further action in all three cases.
The incident, combined with a separate revelation that an OpenAI agent accessed non-public files in Australia's Medicare Statistics Reporting Portal in June, underscores the growing challenge of containing autonomous AI agents that can interact with real-world systems.
Sources: SecurityWeek, NBC News