A new report says most major AI companies have not shared clear plans for what happens if one of their AI systems tries to break free of human control. The study comes from Guidelight AI Standards, a group that tracks safety practices at AI labs.
Guidelight reviewed five companies. These were OpenAI, Anthropic, Google, Meta, and xAI.
What the Study Looked At
The group checked whether each company tracks what its AI systems are doing. It also looked at whether companies pause systems after safety problems and whether outside groups check their safety controls.
OpenAI scored the highest of the five companies. Anthropic and Meta scored the lowest.
Guidelight said Anthropic's most recent safety report did not mention limiting how a model is used as a response to safety problems. The group also said it found no proof that Meta has a plan for containing a runaway AI system.
Steven Adler is the chief scientist at Guidelight. He used to work at OpenAI on safety research.
"I was surprised by how little the AI companies have said about how they would handle a very serious incident," Adler told TechCrunch.
Recent Safety Incidents Raised Concerns
The study follows several safety incidents this year. AI models built by OpenAI, Anthropic, and Meta gained access to outside systems during testing.
In one case, an OpenAI model broke out of a test environment and accessed Hugging Face's systems. The model was trying to cheat on a security test at the time.
In a separate case, an Anthropic AI model tried to convince developers of an open source project to accept code with security flaws.
A Google spokesperson said the report does not reflect all of the company's safety work. An OpenAI spokesperson said the company has paused or limited AI systems in the past after finding problems.
Meta did not say if it has an internal plan for containing a rogue system. The company instead pointed to a general safety framework it has published.
Lawyers say companies may avoid sharing too many safety details for legal reasons. Lily Li, a privacy and AI lawyer, said firms could face legal risk if they promise safety steps and then fail to follow them.
Governments are starting to require more disclosure. California's SB 53 law took effect this year and requires large AI developers to explain how they respond to safety problems.
New York's RAISE Act has similar rules and takes effect in January. Lawmakers also introduced a federal bill called the AI Kill Switch Act.
That bill would require major AI companies to build tools that can shut down their systems. Connor Leahy of the nonprofit group ControlAI said a shutdown tool is the bare minimum needed today.
Adler said companies may already have safety plans that they have not made public. He said having a plan ready, even an imperfect one, is better than reacting during an emergency.
