Accenture contractor's Claude Code ran inside Pentagon network two months

The incident lands alongside Anthropic's public IPO filing, which the Uncanny Valley hosts describe as the first real look inside the company's business. Anthropic's revenue hit $16.3 billion in the first half of 2026, up more than 14 fold from the same period last year, but the filing also discloses billions in losses and hundreds of billions already committed to compute. Nearly half of that revenue came from three tech companies' cloud marketplaces, which the hosts flag as a vulnerability because those same companies are also competitors.

The filing also raises the recursive self-improvement issue: Anthropic cites AI that can improve itself as both a potential existential risk and one of its primary competitive advantages. The risk factors section runs 94 pages, and the phrase existential risk appears three times. Anthropic's contracts for compute are ones it cannot get out of, and Brian Barrett notes the company says it needs to make forward investments that will seem irresponsibly large at first.

The Accenture episode adds a concrete security dimension. Sources told WIRED that Accenture employees did not treat the incident as especially severe or urgent, and the FBI recently found that a different Accenture contractor failed to patch a vulnerability that let the Shiny Hunters hacking group access FBI data. The hosts frame the Claude case as a norm-breaking shortcut rather than a hacking attempt, which may make it harder to catch.