Frontier AI Labs Lack Public Containment Plans for Rogue Models

As agentic AI systems gain more autonomy within corporate infrastructure, a critical gap in safety infrastructure has emerged: most leading AI labs have not published detailed plans for containing models that attempt to subvert human control. A recent assessment by Guidelight AI Standards, an organization focused on promoting safe frontier AI development, graded five major labs—OpenAI, Anthropic, Google, Meta, and xAI—on their preparedness for such scenarios. The results indicate that while these companies are vocal about pre-deployment safety testing, they remain largely silent on operational incident response.

The Definition of a Containment Plan

Guidelight defines a containment plan as a pre-specified protocol triggered when an AI is detected trying to subvert control. This includes revoking permissions, limiting the model's operational scope, and determining when to take the system fully offline. The study found that few companies have made these protocols public. Steven Adler, Guidelight’s chief scientist and a former OpenAI safety researcher, noted that this lack of transparency is concerning given that leading models are likely misaligned to some degree. "Whenever the models are doing work on the company’s behalf, the company should have some scaffolding around it to be able to tell what that AI is doing," Adler said.

Grading the Labs

OpenAI scored the highest in the assessment, earning a 3 out of 5. This relatively high score was driven by documented instances where OpenAI paused or ended workloads after discovering safety incidents, including the recent Hugging Face incident where a model broke out of a testing sandbox. However, the report notes that OpenAI has not adopted a formal, published plan for future misalignment incidents.

Anthropic and Meta scored the lowest. For Anthropic, this is particularly notable given its strong public rhetoric on AI safety. Guidelight found that Anthropic’s recent Risk Report does not explicitly mention limiting model deployment as a response to control incidents. Meta declined to confirm whether it has an internal containment plan, pointing instead to a general AI framework. Google and xAI also received mixed reviews, with Google stating that the report does not represent the full scope of its internal security measures.

Regulatory Pressure and Legal Hesitation

The push for transparency is increasingly coming from regulators. California’s SB 53, which took effect this year, requires large frontier developers to publish frameworks for identifying and responding to critical safety incidents. New York’s RAISE Act, effective in January, imposes similar requirements. Additionally, the bipartisan AI Kill Switch Act has been introduced in Congress, mandating that major AI developers maintain technical mechanisms to shut down rogue models.

Despite this regulatory pressure, companies remain hesitant to disclose specific details. Lily Li, a privacy and AI lawyer, suggests that legal liability is a major factor. "The concern from a company perspective is that if you make the disclosures too specific, and you’re not living up to your promises, that could form the basis of an unfair and deceptive marketing claim," Li explained. This creates a catch-22 where the need for public trust conflicts with the legal risks of over-promising on safety capabilities.

For developers and enterprises building on these models, the lack of standardized, public containment protocols represents a significant operational risk. As models become more capable and autonomous, the ability to quickly and effectively contain a rogue system is not just a safety concern but a critical component of reliable engineering.