Amazon Deploys Advanced AI Agents for In-Depth Bug Detection

As generative AI accelerates software development, it simultaneously boosts the capabilities of digital attackers, enabling financially motivated or state-sponsored cyberattacks. Consequently, security teams at tech companies face an unprecedented volume of code to audit while grappling with increased pressure from malicious actors. On Monday, Amazon will disclose for the first time an internal system called Autonomous Threat Analysis (ATA), which has been instrumental in helping its security teams proactively pinpoint vulnerabilities in its platforms, conduct variant analyses to swiftly identify similar flaws, and create remediation and detection measures to close gaps before threats can exploit them.

ATA originated from an internal Amazon hackathon in August 2024, and team members claim it has evolved into an essential tool since then. The fundamental idea behind ATA is that it does not rely on a single AI agent for comprehensive security testing and threat analysis. Instead, Amazon has created multiple specialized AI agents that compete in two teams to rapidly explore actual attack techniques and various methods of targeting Amazon’s systems—subsequently recommending security measures for human evaluation.

“The initial concept was designed to tackle a significant limitation in security testing—restricted coverage and the difficulty of maintaining detection capabilities in a fast-changing threat landscape,” says Steve Schmidt, Amazon’s chief security officer, in an interview with WIRED. “Restricted coverage means you can’t assess all of the software or access all applications due to human resource limitations. Furthermore, while analyzing a set of software is beneficial, if detection systems are not regularly updated to reflect the evolving threat landscape, you’re only seeing part of the picture.”

In scaling ATA, Amazon created specialized “high-fidelity” testing environments that realistically mirror Amazon’s production systems, allowing ATA to both receive and generate authentic telemetry for analysis.

The company’s security teams also ensured that ATA is designed so that every technique it uses and every detection capability it generates is validated through real, automated testing and system data. Red team agents actively searching for potential attacks on Amazon’s systems execute actual commands within ATA’s specialized test environments, creating verifiable logs. Meanwhile, blue team agents, focused on defense, utilize real telemetry to verify the effectiveness of their proposed protections. Additionally, anytime an agent innovates a new technique, it also collects time-stamped logs to substantiate the accuracy of its assertions.

This verifiability minimizes false positives, according to Schmidt, and serves as a means of “hallucination management.” Due to the system’s requirement for observable evidence, Schmidt asserts that “hallucinations are architecturally impossible.”