The AI agent inside the Department of Defense
In this week's episode of WIRED's Uncanny Valley podcast, Brian Barrett, Leah Feiger, and Zoë Schiffer discussed a scoop of theirs: a defense contractor connected Claude Code, Anthropic's coding product, to a secure system of the U.S. Department of Defense. In the source text, the use is described as one of the most unusual applications of AI in government observed so far.
The material does not clarify what functions the agent performed in the system, how long it remained connected, what authorizations were in force, or whether access has been revoked. These elements remain to be verified: the news, as presented, concerns the existence of the connection and its exceptional nature, not the technical methods or operational consequences.
The IPO filing: extreme numbers in both directions
On Monday, Anthropic publicly filed its application to list on the stock exchange. According to the podcast's account, the company aimed to raise more capital than any other company in history and would be the first major U.S. company dedicated exclusively to developing AI software to go public. The valuation went from $183 billion a little more than a year ago to $965 billion in the latest round, with expectations, reported in the podcast, of approaching $2 trillion at the start of trading.
The document offers the first real look at the accounts of a company that has so far been private. Revenue reached $16.3 billion in the first half of 2026, growing more than fourteenfold compared with the same period the previous year. Losses, however, are also in the billions, and the company has already committed hundreds of billions to computing capacity. Barrett observed that the numbers appeared better than expected and consistent with a real company in growth, though without dispelling all concerns.
Commercial dependencies and declared risks
A critical point raised in the podcast concerns revenue concentration: almost half of the cited revenue would come from three technology companies active in cloud marketplaces, which are also competitors of Anthropic. Barrett called this a cautionary sign, questioning the durability of the competitive advantage in the face of rival AI products and open-source or open-weight models. Schiffer recalled a Ramp estimate according to which about 80% of Anthropic's and OpenAI's revenue would come from roughly the 1% of customers who spend the most, calling the dependency a vulnerability in the event of reduced spending.
In the filing, according to the podcast, the company explicitly stated that sufficiently advanced AI systems, if not developed and protected adequately, could pose catastrophic or existential risks to humanity; the expression recurs several times. The hosts read this transparency as a positive element, while stressing that it is a position long known for Anthropic and not a novelty. It remains to be verified, with the complete documentation, the exact scope of the clauses on risk factors and the detail of the revenue concentration.
In the background, the midterm elections
The podcast also addressed a third topic, unrelated to Anthropic's accounts: volunteer groups across the country are reportedly preparing for possible ICE agent presence at polling stations in the midterm elections, now imminent. The source presents the initiative as a reflection of a deeply uncertain electoral cycle, without providing data on numbers or organizations involved.
