Claude Mythos Model Removed from Market on June 21, NSA Still Has Access Sparks Double Standard Controversy

On June 21, 2026, Anthropic's Claude Mythos model was restricted or withdrawn from the market due to government cybersecurity concerns, while agencies such as the NSA can still use it.

Actual Discrepancies in Policy Implementation

Official notices show that API access to the Mythos model was cut off on June 21, preventing ordinary enterprises and research institutions from continuing to call it. The model had previously passed multiple rounds of security tests, with training data reaching hundreds of billions of tokens. The withdrawal decision directly impacts commercial application development that relied on the model, but the NSA's internal systems retain full access. The government has not disclosed specific technical vulnerability details, citing only "potential supply chain risks."

Internal Conflicts in the Regulatory Framework

The current U.S. AI export controls set licensing thresholds for models with specific parameter counts, and the Mythos model's parameter scale falls within a sensitive range. The NSA, as an intelligence agency, received an exception authorization based on the cybersecurity executive order updated in 2018. Ordinary users are subject to new Commerce Department regulations. The parallel operation of two sets of rules results in different usage permissions for the same model. Bipartisan lawmakers issued statements on June 22, pointing out that such exceptions may weaken overall compliance effectiveness.

Security reviews must cover all users, otherwise the rules are meaningless. — Statement from Senator Maria Cantwell's office

Market Competition and Technology Allocation

After the withdrawal, the originally planned commercial API pricing for the Mythos model, ranging from $15 to $30 per million tokens, was forced to be suspended. Competitors' similar models are still available normally, and some users have switched to other platforms. Industry data shows that the Mythos model previously achieved a pass rate 8 percentage points higher than similar models in code generation and long text reasoning tasks. The restrictions objectively changed the market share distribution.

Analysis of Deeper Driving Factors

The core of the incident lies in the priority ordering between advanced model training resources and intelligence agency needs. The Mythos model uses specific encryption protocols and distributed training architecture, features that may be seen as a double-edged sword: enhancing anti-attack capabilities while increasing potential leakage paths. The government's decision to retain NSA access reflects a focus on the model's actual operational capabilities rather than mere parameter scale control. Opposition voices center on insufficient procedural transparency; the existing approval process does not require public disclosure of quantitative evaluation criteria for exception authorizations.

From a technical execution perspective, after the model's withdrawal, Anthropic did not provide migration tools or alternative solutions. Some internal systems that integrated the model face reconstruction pressure. The NSA's retained usage means it already has independent operation and maintenance capabilities, including model fine-tuning and security hardening. Ordinary users need to wait for new version compliance reviews, which may take several months.

Independent Judgment

This handling shows that U.S. AI regulation is shifting from parameter thresholds to actual deployment controls. The existence of exception authorizations weakens the universal applicability of the rules and may, in the long term, reduce enterprises' trust in compliance paths. The model's technical implementation capabilities have not been denied; the problem lies in the clarity of the allocation mechanism.