LLM News

Researchers have proposed MAGE, a meta-reinforcement learning framework that enables large language model agents to adapt and strategize in dynamic environments. Unlike existing approaches that struggle with long-term adaptation, MAGE embeds the learning process directly within the model by integrating interaction histories and reflections into the context window.

A new AI agent framework called ARTEMIS discovered 9 valid vulnerabilities in live penetration testing against a university network with ~8,000 hosts, outperforming 9 of 10 human cybersecurity professionals. The system achieved an 82% valid submission rate and costs $18/hour compared to $60/hour for professional penetration testers, though it struggles with GUI-based tasks and produces higher false-positive rates.