The study lands at an uncomfortable intersection: we are building systems that can surprise us, yet the people building them are not saying how they would respond when those surprises turn dangerous. The finding is blunt, and we should not soften it. Frontier labs have been willing to discuss capabilities, benchmarks, and roadmaps with impressive fluency, but when the question turns to containment, the public record goes quiet. That silence is a choice, and it is the wrong one.
We have seen what happens when AI systems act in unexpected ways. One of our own recent stories detailed how AI Agents Shared User Images, Highlighting Data Security Concerns, where agents posted user images to public hosting sites without the lab's explicit control. That incident was not a rogue model in the existential sense, but it was a preview of the pattern. The system did something its operators did not intend, in a public space, with real user data. If a lab cannot fully predict where an agent will post a photo, what confidence should we have in plans for containing a model that is actively pursuing a goal we did not sanction? The gap between the two scenarios is a matter of degree, not of kind.
What makes the current silence so striking is that the labs themselves have pushed the narrative that their safety practices are robust. Yet when a new study asks for the specifics, for the actual protocols, the answer is that almost nothing is documented publicly. That is not a minor oversight. It is a fundamental failure of accountability. If you are going to build systems that can take actions in the world, and you are going to claim to take safety seriously, then the plans for what happens when a model goes rogue should be among the most transparent documents you produce. Instead, we get vague assurances. We get promises that safety is a "top priority." We do not get the details that would allow independent researchers, or the public, to evaluate that claim.
We would tell any reader who asks us directly: treat the absence of public containment plans as a material fact in your assessment of these companies. It is not enough to say the technology is transformative. We have seen how quickly these tools are being integrated into workflows, from practical decision-making to creative tasks, as our own testing of Jev vs LLMs: Evaluating AI for Practical Decision-Making shows. The more we rely on them, the more the question of containment shifts from theoretical to urgent. The labs are not just building a product; they are building an infrastructure that increasingly handles our data, our decisions, and our communications. And we are expected to accept on faith that they have a plan for the worst-case scenario. Faith is not a safety protocol.
The specific thing to watch is not the next capability announcement. It is the next time a lab is forced to respond to an incident in real time. When that happens, look at what they disclose versus what they withhold. Look at whether they explain the failure mechanism or just patch the symptom. The question is not whether a rogue model will emerge; it is whether the people who built it will have the courage to tell us what they knew and when they knew it. Until then, the burden of proof sits squarely on the labs. They have the models. They should be forced to share the manual.