How Amazon Predicts and Responds to Cyber Threats with AI Agents

Autonomous Threat Analysis (ATA) is a security testing system that started as an internal hackathon project at Amazon, combining AI agents with adversarial multiagent reinforcement learning to automatically and repeatedly validate attack and defense.

The initial prototype, started in August 2024, was built in just 48 hours and proved its potential by finding gaps in existing detection rules and automatically generating better solutions. ATA has since evolved into a formal security testing framework where red-team and blue-team agents operate together.

The core of ATA is a graph workflow system. Each node is a specialized AI agent with a different role and goal, and the output of one agent feeds into the input of the next.

• red-team agent: mimics attacker techniques and executes actual commands in the test system
• blue-team agent: validates detection coverage and creates or improves new rules
• isolated environment: runs in an environment that resembles real codebases and production systems but is completely separated

In particular, through the grounded execution architecture, the system doesn't just trust the AI's reasoning alone—every technique and detection is validated against real infrastructure and logs. Attack attempts are recorded as timestamped logs and host records, and the blue team checks precision and recall by querying the actual log database.

In the Python reverse shell test case presented as an example, the red team successfully executed 37 technique variants, and in the subsequent focused testing, detection rules were improved based on 64 variants, achieving 1.00 precision and 1.00 recall. These results were reproduced across multiple independent runs, leading to additional threat-hunting opportunities and new detection rules.

The operational impact was also significant. For failed techniques, agents analyzed the errors and self-corrected, typically succeeding within 3 attempts, and the entire workflow was reduced from weeks of manual work to approximately 4 hours, a 96% reduction. In one multi-step test, the system simulated an entire flow including reconnaissance, exploitation, and lateral movement, finding 2 new detection opportunities in under 1 hour.

ATA can run 10 to 30 technique variants in parallel at once, and testing an individual detection rule takes 1 to 3 hours depending on scope and level of parallelization. Through this, Amazon is building a structure where AI handles repetitive security testing while people focus on strategy and contextual judgment.