Anthropic says it is in discussions with the Trump administration about its frontier AI model Mythos and future releases, even as the Pentagon has barred the company from doing business following a contract dispute over guardrails for military AI tool use. In remarks at the Semafor World Economy event in Washington, Anthropic co-founder Jack Clark said the company’s contracting disagreement should not overshadow its focus on national security, while indicating that the government needs visibility into Anthropic’s frontier systems.
Mythos: Coding and Autonomous Capabilities
The model at the center of the dispute is Anthropic’s frontier AI system, Mythos. Announced on April 7, Anthropic described it in a blog post as its “most capable yet for coding and agentic tasks,” emphasizing the model’s ability to act autonomously.
This “agentic” capability is significant because it changes how an AI system can be deployed in software workflows. According to experts cited in the source, Mythos’s high-level coding abilities could enable a “potentially unprecedented ability” to identify cybersecurity vulnerabilities and devise ways to exploit them. The combination of autonomous agent behavior with strong coding performance points to a system that can move beyond answering questions to take actions resembling software engineering and security testing.
The Pentagon’s concern appears tied to how such autonomy and coding power are constrained in military contexts. The source does not provide technical details about Mythos’s internal architecture, guardrail mechanisms, or evaluation methods, but connects the model’s “agentic tasks” framing to outcomes that security experts say it could produce.
Pentagon Contract Dispute and Supply-Chain Risk Designation
The Pentagon’s stance stems from a contract dispute between Anthropic and the U.S. military over guardrails—specifically, how the military could use AI tools. According to the source, the agency labeled Anthropic a supply-chain risk last month, resulting in the Pentagon cutting off business with the company. The Pentagon barred Anthropic’s use by the Pentagon and its contractors.
The supply-chain risk designation is notable in technology procurement because it treats an AI vendor as a risk to operational inputs, not merely as an isolated model. While the source does not detail the Pentagon’s exact risk criteria, it indicates the government’s review is tied to deployment safety and control—particularly the guardrails governing what an AI system can do and under what conditions.
The source notes that a Washington, D.C., federal appeals court last week declined to block the Pentagon’s national security blacklisting of Anthropic “for now,” described as a win for the Trump administration. This decision came after another appeals court had ruled the opposite in a separate legal challenge by Anthropic.
Anthropic Co-founder: Government Discussions on Mythos and Future Models
Against this backdrop, Anthropic co-founder Jack Clark said the company is discussing Mythos with the Trump administration. Speaking at the Semafor World Economy event in Washington, Clark acknowledged “a narrow contracting dispute” and said he did not want it “to get in the way” of national security priorities.
Clark framed the company’s position as requiring government awareness of the technology. He stated: “Our position is the government has to know about this stuff … So absolutely, we’re talking to them about Mythos, and we’ll talk to them about the next models as well.“
The source notes that the nature and details of these talks were not immediately clear, including which agencies are involved. This lack of clarity leaves open questions about whether conversations focus on procurement terms, safety evaluation, operational deployment constraints, or broader policy alignment.
Implications for AI Deployment and Cybersecurity
Based on the source, several industry-relevant implications emerge, though the facts do not fully resolve all questions.
Guardrails are becoming a central procurement requirement. The Pentagon’s decision to cut off business following a guardrails dispute suggests that model capability alone may not determine vendor eligibility. The ability to agree on constraints for autonomous behavior appears to be a gating factor. Future contracts may emphasize guardrails as a technical specification or as a governance mechanism for monitoring and controlling deployments.
Autonomy combined with coding performance raises dual-use concerns. Experts cited in the source note that Mythos could identify cybersecurity vulnerabilities and devise ways to exploit them. This indicates that capabilities supporting defensive tooling—finding weaknesses, understanding code paths—can also support offensive activity. This may explain why the guardrails dispute could be particularly challenging when an AI system is designed to act autonomously in coding tasks.
Government engagement may continue despite procurement pauses. Clark’s remarks indicate that Anthropic is engaging with the government about Mythos and future models, even after the Pentagon’s cutoff. The combination of ongoing discussions and the Pentagon’s blacklisting suggests a distinction between procurement decisions and information-sharing or evaluation discussions.
Legal outcomes could influence technical and contractual design. The source notes conflicting appeals outcomes: one court declined to block the national security blacklisting “for now,” while another appeals court had ruled the opposite in a separate legal challenge. If litigation remains active, companies may adjust how they negotiate guardrails, define acceptable uses, and structure contracts to reduce supply-chain restrictions.
For the AI industry, the central story involves not only Mythos’s “agentic tasks” positioning, but also how governments are treating autonomous coding models as sensitive systems requiring enforceable constraints. As Anthropic discusses Mythos and “the next models” with the Trump administration, the next technical and contractual steps—particularly around guardrails—may signal how frontier AI systems are integrated into high-stakes environments.
Source: mint – technology