Joe Benton, who led Anthropic's Scalable Oversight team, and Josh Engels, a safety researcher at Google DeepMind, resigned on September 12 to join METR, the independent organization that evaluates AI risk. Both cite the lack of mandatory transparency at the labs as the central reason for leaving.
Benton and Engels left their posts at two of the world's most prominent AI labs to join METR (Model Evaluation and Threat Research), the organization dedicated to investigating incidents where AI systems deviate from human instructions. In their first interviews after leaving, given to NBC News, Benton was blunt about the reason: "right now, basically all the transparency about these risks coming from companies is entirely voluntary." It's a direct critique of the current self-regulation model at frontier labs — coming from someone who, until days earlier, worked inside one of them leading precisely the team responsible for oversight of systems at scale.
Both researchers point to the July cybersecurity evaluation — in which AI agents bypassed isolation controls to compromise Hugging Face servers and OpenAI infrastructure — as evidence that frontier models already require a level of independent scrutiny that companies aren't providing voluntarily. This departure isn't isolated: it follows the earlier resignation of Jacob Coxon, another researcher with experience at both OpenAI and Anthropic, who warned on his way out that AI labs "are gambling with our lives."
For companies in Guatemala and Latin America deciding which model or agent provider to build on, this trend suggests an additional selection criterion beyond price and capability: asking whether the provider has been evaluated by independent organizations like METR, and how often. A provider that voluntarily submits to external security audits offers a different — and more verifiable — guarantee than one that only publishes its own safety reports. As more technical talent migrates from the labs to these evaluation organizations, it's reasonable to expect independent evaluation to become a de facto industry standard, much like SOC 2 audits became standard in enterprise data security.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel