The Harbor evaluation tool can now run each eval task inside Docker Sandboxes, microVM environments built to run agents such as Claude Code, Codex and Cursor safely. Custom network policies help keep evaluations fair and repeatable, either locally or in the cloud.
At AI Engineer New York, frontier labs and financial firms discussed how much autonomy to give AI agents in finance. The focus was on keeping consequential decisions under control while still letting agents do useful work.
Docker outlines the main security risks of autonomous AI agents, including prompt injection, excessive autonomy and exposed credentials. The guide also covers ways to contain those risks.
In a Latent Space interview, Leash Bio's Ian Quigley explains how the company mastered DNA-encoded library assays, a notoriously difficult technique, and used the resulting data to detect models that cheat. He also argues the same data is well suited to training AI agents to reason like chemists.
The 3.2.0 release of LangGraph's checkpoint-postgres package adds a prune function to PostgresSaver and AsyncPostgresSaver. It gives developers of agent applications a way to clean up stored checkpoints in their Postgres database.
Hermes Agent 0.22.1 is a patch release with fixes for the Windows Store version and plugin updates. It also adds default parameters for Nemotron Super 3.5, a curated list at the top of the model picker and a Korean language pack for the desktop app.