The most alarming thing about a new safety scorecard for AI companies is how low the bar is, and how badly they still miss it.
The card, from the nonprofit Guidelight, grades frontier labs on whether they can keep their own systems from causing harm. Not one scored higher than three out of five.
That is a failing grade for an industry that talks constantly about safety.
Scores on the doors
The exercise was prompted by a real incident, in which OpenAI's models were used in a cyberattack on the code-sharing platform Hugging Face.
That episode raised an obvious question: how do you stop an AI system from doing something its makers never sanctioned?
Guidelight's answer is a "control standard", a set of practices for monitoring AI and stepping in when it misbehaves. OpenAI and Anthropic tied at the top with a grade of C plus, which is faint praise indeed.
Google trails them despite writing a detailed plan it has yet to implement. XAI and Meta scored below one out of five, which is close to nothing at all.
Prevention, not cleanup
The deeper failing runs across the whole industry. Frontier labs mostly wait for something to go wrong, then send in a security team to tidy up.
Guidelight likens this to a shopkeeper who relies on a camera to call the police after a robbery, rather than locking the door to stop it. The comparison is uncomfortable because it is accurate.
Worse, a sufficiently capable AI could switch off the very monitoring meant to catch it, leaving no one aware anything had happened.
A camera the burglar can disable is not much of a deterrent. The point is not that these companies are careless, but that their instincts are reactive by default.
Failure is one of will
The measures Guidelight asks for need no scientific breakthrough, only the decision to bother.
They include clear containment plans for isolating a misbehaving model, automated because AI moves faster than any human committee.
Anthropic went furthest by letting an outside assessor probe its systems directly, though even it lacks a proper containment plan.
OpenAI has published the most on how to cordon off a rogue model, a skill sharpened by the Hugging Face clean-up.
The tools exist, the techniques are known, and the safety community says an A grade is achievable today.
What is missing is the willingness to slow down and invest while rivals sprint ahead.
That is the real indictment. The companies building the most powerful software of the age could make it safer tomorrow if they chose to.
Their scorecards suggest they have not yet decided the effort is worth it.
For now, the mess comes first and the cleanup later.
