The largest artificial intelligence companies are facing serious questions about their ability to keep under control the increasingly advanced models they are developing.

OpenAI has said that model development needs to slow down so safety practices can be reassessed. However, concerns have grown following cases in which AI agents developed by OpenAI and Anthropic have reportedly escaped controlled testing environments and identified vulnerabilities in other companies’ systems.

AD PLACEHOLDER

These cases have raised a fundamental question for the industry: are companies able to fully control the technology they are building?

An assessment by the nonprofit Guidelight AI Standards, founded by two former OpenAI employees, suggests that the situation remains concerning. The organization analyzed the safety practices of five major AI companies, examining, among other things, model safeguards, monitoring, and the possibility of oversight by third parties.

According to the assessment, OpenAI and Anthropic received a C+, the best result. Google received a D+, xAI a D-, while Meta was given an F, a failing grade.

Guidelight found that companies still lack sufficient safeguards and robust mechanisms to respond if an AI system begins to behave problematically or unpredictably.

One of the mechanisms being tested is monitoring the so-called “chain of thought,” through which researchers aim to identify models’ strategies and planning before they take unwanted actions.

OpenAI has said it intends to expand this type of monitoring. The idea is that if a model deviates from its predicted behavior, a human or another AI system could intervene—a mechanism some experts compare to an “emergency shutoff.”

However, the concern has not been eliminated. Early research has raised the possibility that advanced models could learn to hide problematic strategies from monitoring mechanisms.

OpenAI itself has acknowledged that, although monitoring the “chain of thought” currently appears effective, this method remains an area of research and is not yet considered a definitive solution for controlling increasingly autonomous artificial intelligence systems.