Anthropic AI Updates: September 24, 2026

1. Claude Found an Uncharacterized Enzyme System With CRISPR-Like DNA Repeats

Anthropic. Anthropic introduced a life sciences research group and wet lab formed in spring 2026, and published its first result: Claude, given only high-level direction from scientists, identified a novel enzyme system associated with an array of DNA repeats. The combination of characteristics it found together has only ever appeared in a handful of known systems, all of which are programmable and cut, copy, or paste DNA. The function of this one is not yet known. Source

2. Humans Are Still Gating Every Step in the Lab

Anthropic. TechCrunch’s read of the announcement is that the more notable disclosure is what Anthropic did not do: Claude is not running loose in the biology lab, and scientists remain in the loop at each stage of hypothesis generation and experiment. The workflow is agent-assisted discovery with human approval, not autonomous science. Source

3. An Anthropic Engineer Blamed “Claudeish” Prose on Reward Structure, Not Capability

Anthropic. Jackson Kernion, who works on Claude fine-tuning, explained publicly that newer Claude models write worse because reinforcement learning rewards optimized for math, code, and technical explanation pull the model toward output tuned for other models rather than people. He described the result as writing “adapted to LLM psychology” that reads to humans as overly-dense info dumps, and said countering it requires actively rewarding simple, human-followable explanations. Kernion called Opus 5.5 a better balance while maintaining Opus 4.6 remains his favorite pure writing model. Source

4. Claude Code 2.1.281 Lets the Apps Gateway Assume an IAM Role Into Another AWS Account

Anthropic. The Claude apps gateway can now call Bedrock as an IAM role it assumes through STS, optionally in a different AWS account and optionally one session per developer, and a guardrail: {id, version} setting applies an Amazon Bedrock guardrail to every request through an upstream. Enterprises routing Claude Desktop traffic through their own gateway get cross-account isolation and centralized guardrail enforcement without touching client config. Source

5. Attribution Lines Can Now Be Switched Off in settings.json

Anthropic. Setting "attribution": false in settings.json hides all commit and PR attribution. The release note warns that older CLI versions skip a settings file containing the boolean form, so shared settings files should keep the object form. Source

6. plugin validate Now Catches MCP Config That Would Be Silently Dropped

Anthropic. claude plugin validate gained MCP server checks that report .mcp.json entries that would be dropped at load time, undeclared ${user_config.*} references, and insecure URLs. The same release added browser-based MCP elicitation on 2026-07-28 protocol connections and an auto mode recommendation in /insights that estimates how many permission prompts auto mode would have handled in recent sessions. Source