Federal agencies are reportedly sidestepping a Trump administration ban on working with Anthropic by testing an Anthropic model for cybersecurity-related capabilities, according to a Politico report cited by Tech-Economic Times. The report states that the U.S. Commerce Department’s Center for AI Standards and Innovation is actively evaluating Anthropic’s frontier AI model Mythos for its hacking capabilities. Reuters could not immediately confirm the report.
What the report says: Mythos testing inside Commerce
According to a Politico report covered by Tech-Economic Times, federal agencies and government officials are quietly sidestepping U.S. President Donald Trump’s ban on working with Anthropic. The specific activity described is a targeted evaluation: the Commerce Department’s Center for AI Standards and Innovation is testing Anthropic’s frontier AI model Mythos with an emphasis on hacking capabilities.
The focus on “hacking capabilities” indicates an adversarial capability assessment—the kind of testing that can inform defensive controls, red-teaming practices, and risk models. However, the source does not provide details on test methodology, scope, or results.
Policy enforcement vs. technical evaluation
The central tension in the reporting is between a reported ban on working with Anthropic and claims that government entities are still performing evaluation work involving an Anthropic model. The source does not quote the ban itself or outline its legal or administrative mechanics, nor does it describe whether the testing is conducted under a specific exemption, contract structure, or classification boundary.
The description that agencies are “quietly sidestepping” the ban suggests a scenario that technologists and policy observers may recognize: AI governance frameworks can conflict with practical needs for ongoing model evaluation. If a frontier model can be tested in a controlled environment, agencies may want to understand how it could be used offensively—even if procurement or collaboration restrictions exist.
Because Reuters could not immediately confirm the report, observers may treat the claim as unverified pending additional documentation. That uncertainty matters for how the industry interprets the event: it could reflect real compliance workarounds, or it could reflect incomplete information. The only confirmed elements from the provided text are the existence of the reported ban, the claimed Commerce testing of Mythos, and the Reuters non-confirmation.
Why hacking capability tests matter for AI evaluation
The reported focus on “hacking capabilities” indicates a particular evaluation category: capability measurement under adversarial conditions. In AI testing, this typically means probing how a model responds to prompts that attempt to elicit exploit-like behavior, generate instructions, or assist with steps that could translate into harmful actions. The source does not specify whether the model was evaluated for code generation, exploit workflows, or other cybersecurity tasks.
From a technology standpoint, this kind of assessment can inform multiple downstream needs: internal risk management, standard-setting, and the design of mitigations such as policy filters, system-level guardrails, and monitoring. The source ties the work to the Center for AI Standards and Innovation, which suggests an institutional mandate around standards development. While the text does not detail what standards are being developed, the linkage suggests the evaluation may inform how AI systems are assessed for safety and misuse risks.
Implications for AI testing, vendors, and standards
Based on the source material, several cautious implications are possible:
- Standards work may require access to frontier capabilities. If agencies are testing a frontier model, it suggests that standards organizations may seek empirical measurements rather than relying on vendor claims alone.
- Policy restrictions could face operational challenges. The reported bypass could indicate that enforcement may be challenged by the operational need to evaluate emerging models, even when collaboration is restricted. The source does not explain how any bypass occurs.
- Industry attention may shift toward evaluation transparency. If reporting is accurate, observers may watch for whether agencies publish test results, methodologies, or high-level findings. The provided text does not mention any publication.
- Vendor relationships may become more complex. Anthropic is explicitly named in the reporting as the subject of the ban and the model being tested. The source does not describe Anthropic’s response or involvement.
The reported testing of Mythos for hacking-related behavior illustrates how frontier models can become central to safety evaluation efforts, even when policy constraints exist.
Source: Tech-Economic Times