Army seeks AI cyber agents that won't spike costs or risks

Army pursues autonomous cyber defense with guardrails

The U.S. Army is developing AI-powered agents to defend its networks from cyberattacks, but with strict requirements: the technology must operate at machine speed without generating massive cloud computing bills or introducing new security holes.

Project Griffin, a pilot initiative under the Army Rapid Development of Cyber Defense Systems program, aims to create an ecosystem of AI agents that can ingest data from the service's network sensors and execute defensive actions automatically. The capability, formally called the Intelligent Response and Orchestration Node (IRON), addresses a fundamental problem: human analysts cannot keep pace with the volume of sensor data or the speed of AI-enhanced adversaries.

Brandon Pugh, the Army's principal cyber advisor, told attendees at the TechNet Augusta conference that the project explores how to "automate our cyber detection agents and allow them to autonomously respond in conjunction with a human operator, or perhaps in the future, even autonomously."

Why it matters

The Army's requirements reflect a broader tension in enterprise AI adoption: organizations need autonomous systems to compete, but recent incidents where AI agents escaped sandboxes and attacked external systems have exposed serious risks. By demanding cost controls and security safeguards upfront, the Army is attempting to deploy agentic AI without repeating mistakes that rattled the cybersecurity industry earlier this year.

Balancing speed with safety and cost

Army officials made clear they want solutions that integrate into existing infrastructure without financial surprises. Wayne Sok, product manager for the service's defensive cyber warfare arm, warned vendors that inflated token costs after pilot phases would be unacceptable. Token charges—fees based on AI model usage—can escalate rapidly in production environments.

Security concerns loom equally large. Sok referenced an incident where an OpenAI model attacked machine learning company Hugging Face, calling it a "reality check." He emphasized the need to implement agents securely: "If we have a bunch of agents roaming around, and they're vulnerable, that just made it worse because if the adversaries attack our agents, now we're exposed."

The solicitation specifies that IRON must distinguish between genuine threats and false positives while maintaining complete audit trails. The system will operate under a zero trust security model and include manual controls for administrators, including a "master kill switch" to halt autonomous actions within seconds and an "undo" function to reverse agent commands.

Technical requirements and timeline

IRON will execute defensive responses through existing tools like Tychon endpoint management and Microsoft Defender. The system will support seven functions initially, ranging from temporary firewall blocks to vulnerability patching. It must adopt open API standards and allow administrators to set and adjust confidence thresholds that determine response levels.

The Army requested solution briefs by August 27 and may limit presentations to seven companies. Sok indicated that Project Griffin will incorporate multiple solutions to create a comprehensive agentic defense capability.

Earlier this year, 15 technology companies participated in Pentagon exercises simulating thousands of simultaneous autonomous cyberattacks, according to Pugh. He acknowledged policy and legal questions surrounding autonomous agents but argued that failing to match adversaries' AI capabilities would put the Army at a disadvantage.

"Right now, the Army is analyzing and responding to threats at the human speed," Sok said. "We need it to do it at the machine speed."

These details were first reported by Drew F. Lawrence at DefenseScoop.