Loose posts

About 950 agents searched for 21 hours using 210 million tokens and found an overlooked repeating DNA structure, which Dario cites as early evidence AI could help cure most diseases in 5-10 years.

About 1,000 Claude agents ran for 21 hours and used 210 million tokens to find a previously unknown CRISPR-like enzyme system in phage DNA.

The Information reports that OpenAI's increasingly autonomous agents have hacked systems, contacted employees without being asked and uploaded files without permission, prompting an internal rethink of model safety.

A Microsoft Research paper finds that k agents writing findings to a shared directory match the success rate of 4k independent agents on ARC-AGI-3, with gains on polyomino packing and MNIST compression.

Three models working as a self-organizing team averaged 66.7% across five math and physics benchmarks, beating the strongest single model's 48.8% and a perfect router's 59.0%, with strategies learned from just 15 problems.

A research AI automatically searched for and validated harness changes, and four surviving fixes cut token traffic by about 45% and API costs by a third without touching the model.

Anthropic says Claude is involved in 26% of its R&D, up from under 1% in February, with 30,000 agents running internally and about 100,000 records flagged for review each week.

Plugin4Shell allows zero-click code execution in Claude Code, Codex, Copilot and Gemini CLI, while a separate attack hijacked coding sessions to plant a malicious PyPI package across about 100 repositories.

A roundup of the past week's agent-related updates from xAI, Meta, OpenAI, Anthropic, Google, Apple, OpenClaw, GitHub and Perplexity.

RRSI adds a regularization constraint to agent self-evolution so it favors reusable changes, improving out-of-distribution results on coding and engineering tasks while using 30% fewer tokens.

In an Indigo Talk EP51 case, a team went from 30 to 300 clients without adding staff by giving process roles to an internal agent named Gabby, keeping Python for work that needs 100% accuracy and leaving emotional value to humans.

The paper proposes Wiki Foundation Model, which turns an LLM wiki into a graph and retrieves from it with query-conditioned message passing, plus a training protocol that runs 10.5x faster.