- OpenAI added Record & Replay to Codex on macOS. Users can demonstrate a workflow once — such as uploading a YouTube video with metadata, thumbnail, and subtitles — and Codex turns it into a reusable “skill” it can run again. The update also adds local-to-remote handoff, pushing Codex further from coding assistant toward repeatable work automation.
- AWS framed production AI agents as a context and security problem. AWS Context builds a knowledge graph from enterprise data so agents can reason with business context and access controls, while AWS Continuum automates vulnerability detection, prioritization, validation, and remediation suggestions.
- Hugging Face’s MosaicLeaks shows how research agents can leak private information through search queries. Even if no single query exposes a secret, an observer can reconstruct sensitive facts by combining query fragments. Simple “don’t leak” prompting helped inconsistently, and training only for task performance made leakage worse because agents packed more private context into web searches.
- Hugging Face proposed a more practical way to benchmark open models as agents. Instead of checking only the final answer, the benchmark looks at effort: tokens, time, turns, tool calls, and whether the model follows clean APIs or works around them. The point is simple: two agents can get the same answer, but one may do it with a single CLI call while another burns time debugging a custom script.
- Anthropic hired John Jumper from Google DeepMind. Jumper shared the 2024 Nobel Prize in chemistry for AlphaFold and was reportedly involved in Google’s coding-tools efforts. The move highlights how aggressively frontier labs are competing for scientific and agentic-coding talent.
- The U.S. government’s Anthropic crackdown is turning into a policy, export-control, and branding story. TechCrunch discusses why Anthropic took Fable 5 and Mythos 5 offline, why cybersecurity experts argue the order could hurt defenders, and how the ban may paradoxically make Anthropic’s models look more powerful.
- A UC Berkeley study suggests ChatGPT-era grade inflation is driven more by outsourced work than better learning. In courses heavy on writing and coding, grades rose sharply after ChatGPT launched, especially where homework carried more weight. That points to a growing need for assessments that verify process and understanding, not just finished output.
Bottom line: as agents take on more real work, context, security, evaluation, and auditability are moving from “nice to have” into the core of AI products.