- A reported June 2, 2026, US executive order classified frontier AI model evaluations following an alleged zero-day vulnerability discovery involving Anthropic’s Claude Mythos model, according to reports.
- Classified evaluation results cut off independent researchers and regulators from the evidence base needed to assess vendor safety claims or build shared benchmarks, with no compensating mechanism for systems already deployed.
- Enterprises deploying high-risk AI systems before the EU AI Act’s August 2, 2026, transparency deadlines cannot rely on post-deployment retrofits to close compliance gaps, risk classification, technical documentation and human oversight procedures must be in place before the deadline, not after.
A reported executive order dated June 2, 2026, classified the results of frontier AI model evaluations, according to reports, after an alleged zero-day vulnerability discovery involving Anthropic‘s Claude Mythos model. The order’s immediate effect is to remove the public evidence base that independent researchers and regulators depend on to assess what frontier models can actually do, at the precise moment the EU AI Act’s high-risk compliance deadlines are approaching.
The Evaluation-Deployment Gap Widens
Before the order, a programme of AI model evaluations was building a public record of frontier model capabilities. That record is now restricted.
The consequences reach beyond the immediate security concern. Many advanced AI systems are already embedded in critical infrastructure and enterprise operations, often without meaningful external scrutiny of their potential for misuse or emergent behaviour. The UK AI Security Institute has observed that evaluation in synthetic environments gives only a limited picture of real-world risk, and that strong benchmark scores do not reliably translate to finding or fixing actual vulnerabilities. The problem is structural: internal red-teaming and benchmark performance are designed to validate capability, not to surface the kind of adversarial behaviour the Mythos incident reportedly exemplified.
Consequences of Reactive Safety Protocols
Classifying evaluation results creates a direct tension with the transparency that shared safety standards require. When findings are restricted, other organisations, researchers and regulators lose access to the evidence base needed to assess vendor claims, build comparable benchmarks or understand what risk profiles actually look like at the frontier. Anthropic, OpenAI and Google DeepMind had previously engaged in more open evaluation work with CAISI. That collaboration now operates under constraints that limit what can be shared publicly, according to reports.
The confidentiality may be necessary given what the Mythos incident reportedly revealed. That does not make the trade-off cost-free. Retrofitting safety mechanisms onto already-active systems is technically complex and resource-intensive, and the classified framework does nothing to address AI systems deployed before the order. The reactive cycle, capability first, evaluation second, classification third, means the gap between what models can do and what governance frameworks can manage tends to widen with each generation. For enterprises building AI-dependent operations, that is a compliance and liability concern as much as a safety one. The EU AI Act’s August 2026 transparency deadlines arrive directly against this backdrop.
The Slow March of AI Regulation
The EU AI Act entered into force on August 1, 2024with most high-risk AI obligations applying from August 2, 2026. For high-risk AI systems embedded in products subject to third-party conformity assessment, the compliance deadline extends to August 2, 2027. That phased structure means many AI systems will have been in widespread deployment for a year or more before the full weight of the regulation applies.
The infrastructure for conducting those assessments is also still developing. The notified bodies that will carry out third-party conformity assessments for high-risk AI systems are at an early stage, and the methodologies for evaluating general-purpose AI models remain contested. Organisations currently deploying AI are in an active compliance window: risk classification, technical documentation and human oversight procedures all need to be in place before the deadlines, not after. The pattern the EU AI Act follows, regulation responding to market adoption rather than preceding it, is consistent with the broader trajectory the US executive order illustrates. For more on the Act’s implementation structure, see our coverage of how the EU AI Act’s four-tier risk framework operates.
Building Proactive, Integrated Safety
The core argument running through this situation is that deployment is where risk lives, that most AI risk stems not from model capability alone, but from how organisations choose to use that capability.
Continuous post-deployment monitoring is the piece current frameworks most consistently underweight. Validation metrics measure performance in controlled conditions; they do not track how a system behaves as environments shift, as users push edge cases, or as the model’s outputs interact with other systems over time. The Mythos incident is an extreme case of emergent capability surfacing at evaluation rather than deployment. The more common failure runs the other way, with problems appearing in production that pre-deployment testing did not anticipate. Embedding safety into model design, training, evaluation and ongoing operation is less a governance aspiration than a practical requirement, one the June 2026 executive order made difficult to ignore. For more on AI agents and automation tools, visit our AI Agents section.
