Major AI research labs are still not sharing publicly how they’d contain a model that acts independently or dangerously—despite guidance pushing them to build “rogue model” response plans for scenarios where AI circumvents human control. A new assessment by standards group Guidelight AI examines exactly that question across OpenAI, Meta, Google, Anthropic, and xAI, finding that only some have any public containment protocols in place.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/))
How Labs Fared in the Containment Study
Guidelight AI graded these five frontier AI developers on metrics such as monitoring systems, stopping models after misbehavior, third-party auditing, and how clearly they’ve laid out trigger points for shutting down or restricting errant models.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/)) The evaluation revealed OpenAI as the most transparent overall—scoring 3 out of 5—thanks to actions such as pausing model workloads when safety incidents occur and describing some steps taken toward regaining control of misbehaving systems.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/))
By contrast, both Meta and Anthropic received low marks for containing rogue models. Anthropic’s public reports do not clearly state that restricting a model (“limiting deployment”) is an outcome of their misalignment or risk responses.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/)) Meta, meanwhile, had no publicly available containment plan and showed no sign of committing to one directly.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/)) Google and xAI also didn’t provide clear public plans, though Google signaled that not all its internal practices are disclosed.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/))
Emerging Pressure from Regulators & Safety Advocates
The lack of public plan clarity comes as regulators in New York and California roll out laws that demand greater transparency from companies building frontier AI. California’s SB-53, effective in 2026, requires disclosure of how AI firms detect, respond to, and manage critical safety failures and risks of systems bypassing oversight.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/)) In New York, the RAISE Act imposes similar obligations starting January.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/)) On the federal side, the proposed AI Kill Switch Act would mandate technical mechanisms to shut down problematic models.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/))
Experts say there are reasons companies may avoid revealing detailed incident response plans. Legal liability, deceptive marketing risks, competitive concerns, or admitting weaknesses publicly could all dissuade full disclosure.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/)) Still, Guidelight and safety researchers emphasize that merely talking about safety is not enough—companies need to prepare pre-specified, actionable plans for when misaligned models threaten safety or autonomy.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/))
For its part, OpenAI maintains it restricts permissions, pauses or limits deployment, or takes models offline when necessary—though there’s no public formal framework detailing exactly what happens or when.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/)) Anthropic says it would evaluate risks if a model tried to evade control; Meta points to its public AI risk framework but didn’t confirm an internal containment plan; Google confirms it has more internal safety measures than publicly disclosed.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/))
This gap between public disclosure and internal safety readiness means that many containment protocols may exist behind closed doors. But without transparency, it’s hard for customers, regulators, and other stakeholders to assess how seriously labs are treating the operational risk of increasingly autonomous systems.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/))
This issue grows more urgent as agentic AI—that is, AI able to act independently in complex environments—is spreading across business operations. As models gain capabilities and integrate with critical infrastructure, failing to plan for worst-case scenarios risks both policy violations and real harm.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/))
Overall, Guidelight’s study finds that while OpenAI has improved its public handling of containment, all labs still fall short of defining clear, formalized emergency response strategies for rogue model incidents. What exists is patchy, rarely detailed, and often undisclosed.([techcrunch.com](https://techcrunch.com/2026/08/22/frontier-ai-labs-still-wont-say-how-theyd-contain-a-rogue-model/))
What this signals is a turning point: legal requirements are coming, and public trust in AI depends not only on what developers promise when it comes to safety, but what they explicitly plan for when control is compromised. The labs that lead the way on transparent containment protocols may gain both regulatory advantage and user confidence.
