Claude on the Pentagon's Network, Altman's Cost Projections, and a ChatGPT Reversal
How this was made Verified AI
Every Intellegix briefing is generated from that day's broadcast and run through automated checks before it publishes — with a human paged on any flag. Here is the trail for this edition.
A report emerged this week that an Accenture consultant deployed a Claude-based AI agent on a Department of Defense network, apparently without the level of formal authorization that sensitive government infrastructure requires. The agent was reportedly performing a routine work task such as document analysis, but its security implications are independent of its intentions. Claude, like all large language model APIs, processes submitted text on Anthropic's servers — meaning any documents fed to the agent were leaving the DOD's controlled environment and traveling to an external commercial service. For unclassified material that represents a significant data governance breach; above that classification level it would constitute a serious violation.
Security researchers emphasized that the issue is systemic rather than specific to Claude. The same exposure would apply to GPT-4, Gemini, or any externally hosted model used by a government contractor. The General Services Administration and the Office of Management and Budget have both been developing AI governance frameworks for federal agencies, but individual contractors deploying commercial AI tools on government networks represent a gap those frameworks have not fully addressed. The incident is expected to accelerate formal guidance from the DOD's Chief Digital and Artificial Intelligence Office or GSA's IT acquisition division.
The story's timing coincided with a Bloomberg report identifying Anthropic as SpaceX's single largest customer — a relationship apparently involving significant compute infrastructure routed through SpaceX's connectivity network at a scale that surprised observers. That dependency maps onto a broader pattern: Anthropic's primary cloud relationship runs through Amazon Web Services, part of the $4 billion investment Amazon made last year. AI-to-AI infrastructure dependencies are deepening across the industry, concentrating risk in ways that are only beginning to be mapped.
OpenAI CEO Sam Altman, appearing on the Cheeky Pint podcast with John Collison, stated that AI computing costs should fall toward $20 billion for the infrastructure needed to run frontier models, noting that a gigawatt of AI capacity currently costs roughly $70 billion over five years. The direction of the projection — falling costs — is widely shared across the industry, though the rate and the floor remain genuinely uncertain given constraints in power generation, cooling infrastructure, and specialized semiconductor supply chains. Separately, OpenAI reversed a recently announced reduction in ChatGPT Pro usage limits for its $200-per-month tier, not only restoring previous allowances but doubling them after swift backlash from paying subscribers — a sign of how commercially critical that power-user constituency is to the company's enterprise sales pipeline. US computer science enrollment, meanwhile, fell for the second consecutive year, raising longer-term questions about the talent pipeline for an industry whose infrastructure ambitions are expanding rapidly.