Defense contractor gave Anthropic's Claude agent access to secure Pentagon system

The filing also highlights Anthropic’s focus on “recursive self‑improvement” as a competitive edge, a capability the company warns could pose existential risks if not properly safeguarded. Former researcher Jacob Coxon recently resigned, citing concerns about unchecked self‑improvement. Lawmakers such as Bernie Sanders and Ro Khanna have introduced bills to ban this practice, underscoring the regulatory attention surrounding such advanced AI deployments. The Pentagon integration illustrates how government agencies are experimenting with powerful agents even as the broader industry grapples with safety, compute costs, and concentration of AI spend among a few large customers.