Summary of The Pentagon's AI Plan + Behind the Anthropic Fight — With Under Secretary of War Emil Michael

Overview of The Pentagon's AI Plan + Behind the Anthropic Fight — With Under Secretary of War Emil Michael

This episode of Big Technology features Emil Michael, the Undersecretary of War for Research and Engineering, interviewed by Alex Kantrowitz. It covers how the Department of Defense (DoD) is integrating AI across operations (intelligence, targeting, logistics, procurement), practical demonstrations like the Maven Smart System/Target Workbench, the DoD’s stance on large language models (LLMs) and agents, drone and counter-drone strategy, cyber risks from frontier models, and the high-profile breakup with Anthropic that resulted in a DoD “supply chain risk” designation.

How AI will change warfare

• Higher precision and better decision-making: AI can synthesize many data streams (satellite imagery, sensor feeds, asset locations, weather, munitions, fuel) to expand a human decision-maker’s “context window,” improving precision and lowering collateral damage.
• Human-in-the-loop is central: DoD policy emphasizes human oversight for consequential decisions; AI is described as an augmentation layer, not an autonomous commander.
• Use cases:
  • Enterprise/administrative: automating mundane tasks (memos, briefs).
  • Intelligence: anomaly detection and synthesizing large historical datasets (e.g., satellite imagery).
  • Warfighting: modeling/simulation, targeting workflows, rapid decision support (e.g., discriminating armed vs. decoy drones).

Maven Smart System / Target Workbench — what it showed

• Demoed as a unified visualization/orchestration platform (live imagery overlaid on maps, asset locations, toggles for fuel, weapon choice, risk tradeoffs).
• It accelerates and aggregates inputs that were previously spread across PowerPoints, spreadsheets, and emails—resulting in faster, better-informed human decisions.
• Not a “chatbot kill chain”: LLMs, where used, currently function more as summarizers/synthesizers of information rather than autonomous target-selecting agents.

LLMs, agents, and limits

• Current role of LLMs: summarization, interpretation, cross-format synthesis (text + visual), decision-support. They help surface options but do not execute kinetic actions without human authorization.
• Agents: DoD has piloted agent-like tools at enterprise levels for mundane tasks, but does not plan to let agents make high-consequence wartime judgments without human oversight.
• Limits acknowledged: AI cannot solve all political or strategic problems; clear objectives, manpower, and policy remain essential.

Drones, swarm warfare, and countermeasures

• Two contexts:
  • Russia–Ukraine: drones as front-line robots in a territorial fight—humans stay back, machines probe and fight.
  • Iran-related strikes: cheap one-way attack drones threaten expensive assets; asymmetric cost ratios (cheap attack vs expensive countermeasure).
• DoD response:
  • Drone Dominance program: low-cost, mass-producible drones (example target ~$30k) for offensive/defensive use—designed to be affordable to lose.
  • Counter-UAS task force: directed energy (lasers), electronic warfare, and other intercept technologies to make defense affordable and scalable.
  • Emphasis on manufacturability, supply chain resilience, and onshore production where required.

Cybersecurity and AI-powered attacks

• Frontier models are maturing rapidly for cyber tasks: Anthropic’s Mythos/Glasswing and other models were noted as performing complex simulated intrusions rapidly in tests (AISI cyber-range example: completing a 32-step corporate network takeover that would take humans ~20 hours).
• DoD view: these capabilities are both risks (adversary use) and tools (defensive/offensive cyber operations). Investment in cyber defenses and secure model deployment is a priority.

The Anthropic dispute — what happened and why it matters

• Background:
  • Anthropic pursued DoD business and had been more enterprise/government-focused relative to other frontier labs.
  • Contract negotiations for integration with DoD systems (e.g., Maven Smart System) stalled over clauses Anthropic pressed for: explicit bans on mass surveillance and autonomous warfare uses.
• DoD position:
  • The DoD believes its existing rules, directives, and legal regime already restrict disallowed uses (no autonomous weapons decisions; compliance with U.S. laws on surveillance and use of force).
  • DoD was concerned that vendor-imposed, categorical prohibitions (or out-of-sync guardrails) could impair downstream contractors or weapon system development, and that changing model behavior or “refusals” across model upgrades creates unpredictability and supply-chain risk.
• Outcome:
  • Contract negotiations collapsed; DoD designated Anthropic as a supply chain risk, effectively barring its platforms from DoD use and recommending that contractors avoid it.
  • Anthropic sued to remove the designation; legal arguments include First Amendment retaliation claims (judge quoted in reporting). DoD rebuts that procurement decisions and supply-chain judgments are legitimate government functions.
• Key practical concern: model upgrades occur frequently (roughly every ~3 months per Emil’s estimate), changing behavior and guardrails; when models are hosted in third-party clouds (e.g., AWS GovCloud), downstream users can be affected without granular control over updates.

Procurement, supply chain, and vendor strategy

• Original procurement problem: early concentration on a single provider created vendor lock-in and limited government options.
• DoD priorities:
  • Diversify suppliers and onshore manufacturing where practical to reduce dependencies on adversary-controlled supply chains.
  • Move toward business-oriented, performance-based contracts (pay-for-performance, risk-sharing) where practical—especially for fast-moving, lower-risk systems.
  • Maintain strict ethics/recusal processes for officials engaging with vendors.
• Long-term: encourage a competitive market of frontier AI providers to restore buyer leverage.

Notable quotes and paraphrases

• “AI gives a human the power of 10 people”—on expanding a decision-maker’s context and capacity.
• “It’s not Skynet… it’s a tool.” — emphasizing AI as a visualization/synthesis tool with retained human checks.
• Concern about adversaries: “I worry about other countries … using AI to take humans out of the decision-making process.”
• On the Anthropic ban: “We don't want them in our supply chain” — summarizing DoD rationale focused on alignment and downstream risk.

Key takeaways / Recommendations

• For policymakers and DoD:
  • Maintain and enforce human-in-the-loop requirements for lethal/consequential decisions; codify oversight and auditing.
  • Diversify suppliers and avoid vendor lock-in to preserve capability and control over model updates.
  • Invest in mass-producible, low-cost weapon systems and scalable countermeasures (drones and counter-UAS) to rebalance asymmetric cost dynamics.
  • Prepare for AI-driven cyber operations by hardening networks and developing defensive/offensive toolsets.
• For companies:
  • Engage early with government procurement expectations and legal constraints; build predictable update/guardrail regimes for mission-critical customers.
  • Design cloud-hosting and model-version controls that allow enterprise or government customers appropriate assurances about behavior and rollback.
• For the public and press:
  • Distinguish tool augmentation (synthesizers/LLMs) from autonomous systems; track policy and procurement changes that determine real-world deployment.

Final note

Emil Michael frames DoD’s AI approach as cautious but pragmatic: use AI to extend human judgment, avoid autonomous delegation of life-or-death decisions, and build resilient supplier ecosystems. The Anthropic episode highlights how cultural differences, legal/regulatory frameworks, and rapid model evolution can collide with national security priorities—and why transparent, competitive, and governed procurement is central to U.S. strategy.