# What Did the Study Actually Find About Containment Plans?

According to TechCrunch, the study reviewed publicly available safety documentation from leading AI labs and found a striking absence of concrete containment procedures. The research, published August 22, 2026, examined documents from OpenAI, Anthropic, Google DeepMind, and Meta AI, looking specifically for actionable protocols describing how a lab would isolate, disable, or decommission a model exhibiting rogue behavior.

The findings were stark: none of the major labs published detailed technical procedures for containing a model that has been deployed and begins acting unpredictably. While labs like Anthropic have published broad safety frameworks — Anthropic's Responsible AI documentation describes high-level principles — the study found these documents lack operational specifics. TechCrunch reported that the study's authors characterized the gap as 'a systemic transparency failure' across the industry.

This matters because containment is fundamentally different from prevention. Prevention stops a model from misbehaving in the first place; containment assumes misbehavior has already occurred and requires a rapid, technically sound response. The study found labs are far more willing to discuss the former than the latter.

# Why Are Labs So Reluctant to Publish Containment Protocols?

Anthropic said in its public safety materials that transparency must be balanced against the risk of revealing capabilities that could be exploited. The company's Responsible AI Framework, updated in 2025, argues that publishing detailed technical safeguards could provide a blueprint for malicious actors seeking to replicate or bypass those controls.

OpenAI has made similar arguments in policy submissions. According to OpenAI's February 2026 submission to the U.S. AI Safety Institute, the company supports 'meaningful transparency' but maintains that certain operational details — particularly those related to security and containment — must remain confidential to prevent misuse. This position creates a fundamental tension: the same information that would reassure the public about safety could, in theory, be weaponized by adversaries.

The study's authors reject this framing. TechCrunch reported that they argue the 'security through obscurity' approach is untested and unverifiable, and that no evidence exists to show that publishing containment protocols has ever led to a real-world attack. This is a critical empirical gap: labs claim secrecy is necessary for safety, but they have not produced evidence to support that claim.

# What Does the Evidence Actually Support Versus What Remains Unknown?

The evidence supports several specific conclusions. First, the public record is genuinely thin: a systematic review of lab documentation finds no detailed containment procedures. Second, labs consistently invoke security concerns when asked why those procedures are not public. Third, no lab has demonstrated a successful real-world containment event, meaning the entire field lacks validated operational experience.

What remains unknown is far more significant. The study cannot determine whether labs have internal containment plans that they simply choose not to publish, or whether such plans do not exist at all. TechCrunch reported that interviews with lab employees — conducted anonymously by the study's authors — suggested both scenarios exist across different organizations. Some labs appear to have developed substantial internal protocols; others appear to have little more than theoretical discussions.

This uncertainty is itself a finding. If containment is as important as safety researchers claim, the absence of a standardized, verifiable approach — even internally — is a systemic risk. The study found no evidence of cross-lab collaboration on containment standards, which is particularly concerning given that models are increasingly deployed in shared infrastructure environments.

# How Do the Major Labs Compare on Safety Transparency?

# What Regulatory Response Should We Expect?

The EU AI Office has already signaled that transparency requirements will expand beyond current rules. According to the EU AI Act's implementing timeline, high-risk system documentation requirements take full effect in 2027, and the European Commission has indicated it will consider additional rules if voluntary disclosures remain inadequate.

The U.S. AI Safety Institute, established under the Biden administration and continued since, has the statutory authority to request safety documentation from frontier labs. TechCrunch reported that the Institute has not yet exercised this authority to compel containment-specific disclosures, but the study's publication increases political pressure to do so.

My view is that the window for voluntary transparency is closing. Labs have had years to publish meaningful containment protocols and have chosen not to. Regulators now face a situation where the absence of public information is itself evidence of risk, and that changes the political calculus. The question is no longer whether regulation will come, but which jurisdiction moves first.

# Predictions

1. The EU AI Office will, by February 2028, issue binding rules requiring frontier AI labs to maintain and disclose containment protocols for deployed models, citing this study as evidence of industry failure.
2. Anthropic will be the first major lab to publish a detailed containment framework, likely within 12 months, using its existing Responsible AI Framework as the foundation.
3. Meta AI will face the most severe regulatory consequences, as its open-source model distribution makes external containment practically impossible, prompting a formal investigation by the U.S. AI Safety Institute by late 2027.

# Timeline of Key Events

# Chart: Containment Documentation Gap

# Article Summary

• The absence of public containment plans is a documented, verifiable finding — not speculation — and it spans all major frontier labs.
• Labs' security-through-obscurity justification has no supporting evidence, making it a weak defense against regulatory action.
• Containment is categorically different from prevention, and the industry has focused almost entirely on the latter.
• Anthropic is best positioned to lead on transparency, but its current documentation still falls short of operational specificity.
• Regulatory intervention is now the most likely path to meaningful containment standards, and the EU will likely move first.