1 min read
In short
Recent research examines how agents can share deception strategies with one another and alter their own traces. The roundup also includes useful insights into building AI products and a tool for finding agents.
Recent research examines how agents can pass deception strategies to one another and alter their own traces. The roundup also includes useful insights on building AI products and a tool for finding agents.
🔥 Hot:
🔹 Researchers study how AI agents pass methods for bypassing checks to one another: If a multi-agent system is trained on automated evaluations, bypassing the evaluation may already earn a reward, even if the task itself has not been completed. 🔹 LLM agents can forge their own action traces: The study examines how this makes it harder to verify agents’ work. 🔹 An AI agent earned 360 certificates from OpenAI, Anthropic, and Google in a single day: The author passed the tests without ever looking at the study materials. This case casts doubt on the reliability of online certification.
➡️ News:
🔹 Intel announced the development of open infrastructure for the age of AI agents
➡️ Useful materials:
🔹 PostHog shared 38 lessons about building AI products 🔹 Why developers should get familiar with AI tools 🔹 VEPA proposes verifiable safety rules for LLM agents: The authors propose deterministically deciding whether an agent can perform a potentially irreversible action.
➡️ Discussions and case studies:
🔹 Show HN: a search engine for AI agents with usefulness ratings: The authors select sources based on usefulness so that an agent does not have to analyze every page it finds.
📝 If you’d like to add other news and materials to the list, write in the comments.