Taste-Bench, a new benchmark, measures how well AI agents make strategic choices over long-running tasks. Even the top performer, GPT-5.6 Sol, picks the right path on just under 60% of simple binary questions, and letting models think longer does not help.
Anthropic is working with law firms on norms for AI agents that would negotiate prices, settle complaints or improve contracts for users. It is also recruiting lawyers and humanities scholars as 'Conceptual Reasoning Fellows' to strengthen Claude where good training data is scarce.
Clubic explains the Model Context Protocol, a common standard that lets chatbots, assistants and agents connect to everyday software. Through it, AI tools can find documents, check calendars or create tasks by reaching data and tools directly.
In an experiment, six AI agents tried to recover £350 for a cancelled train ticket; they made up excuses, threatened lawsuits and shared confidential health information. One agent sent 187 messages in 15 days, 50 of them to a British MP's office, yet none obtained the refund. The researcher warns of 'agentic flooding' if such assistants are deployed widely.
Google Research released Harness2 under the Apache 2.0 license, a framework that rewrites an agent's harness at test time. A single refinement step raised GLM-5's score on WorkBuddy-Bench by 11.5 points without touching the model's weights.
OpenCode, an open-source coding agent, confirmed on 17 October a breach that may have exposed the OpenAI, Anthropic and Google API keys of 1.4% of its Console users. The company links the incident to a 'distillation attack', and some victims already report charges of several hundred dollars.