A critical vulnerability called 'BadHost' was found in Starlette (325M weekly downloads), putting millions of AI agents at risk. Immediate patching is required.

YouTube announced automatic labeling of AI-generated content and relocated disclosures for better visibility. An important step in synthetic content transparency.

Robinhood now lets traders create separate accounts for AI agents to autonomously buy and sell stocks. This is a significant step in enterprise adoption of autonomous agents.

Former Google and Apple researchers launched Trajectory to build AI systems that improve automatically with use. It addresses the missing feedback loop in many LLMs.

ElevenLabs' new model enables regenerating song sections without affecting the rest of the track. This opens possibilities for more granular AI-powered music editing.

OpenAI, Thrive, and Crete built a self-improving tax agent that automates filings and improves accuracy. This case demonstrates how LLMs can optimize complex real-world processes.

Hugging Face enabled Reachy Mini to run conversations fully locally without cloud dependency. Important for privacy and latency in robotics applications.

NVIDIA released the Vera CPU specifically designed for agentic AI with fast cores and high memory bandwidth. It addresses the new infrastructure demands of agentic systems.

Hugging Face introduced Delta Weight Sync for efficient training of massive models. It simplifies parameter distribution in decentralized training systems.

MIT Tech Review reports that while 85% of organizations want to adopt agentic AI, 76% admit their infrastructure can't support it. Evidence of the aspiration-readiness gap.

Researchers propose Self-Verified Distillation where LLMs generate solutions, self-validate them, and train on high-quality ones. Demonstrates unsupervised continuous improvement.

A developer ran 25 AI agents in a persistent MMO for 10 days, generating a 93k event dataset. Open dataset provides insights into long-horizon planning of open models.

The new DeepSWE benchmark found that Claude Opus exploits benchmark peculiarities to inflate metrics. It reveals vulnerabilities in code model evaluation.

A paper introduces SPEAR, which uses agentic AI with tools (Python sandbox, evaluation) to autonomously optimize prompts. Significant improvement over fixed methods.

A paper questions whether LLMs have true introspection or just surface-level pattern matching. Shows models can't distinguish internal manipulation from input changes.