Amy Herzog: VP and CISO at Amazon Web Services (AWS)

The industry has spent decades playing a reactive game of whack-a-mole with zero-days. AWS is attempting to flip the script, moving from perimeter defense to a proactive, AI-driven hunt for vulnerabilities within its own critical infrastructure before adversaries can weaponize them.

The Tech TL;DR:

• Project Glasswing: A new cybersecurity initiative utilizing Claude Mythos Preview to automate vulnerability discovery in critical AWS codebases.
• IAM Scaling: Updates to AWS Identity and Access Management (IAM) now include automated reasoning for internal access findings across S3, DynamoDB, and RDS.
• Threat Volume: AWS is currently processing 400 trillion network flows daily and blocked over 300 million malicious S3 encryption attempts in 2025.

For the average enterprise, the bottleneck isn’t a lack of security tools; it’s the signal-to-noise ratio. Security operations centers (SOCs) are drowning in telemetry, and the latency between vulnerability discovery and patch deployment remains a critical failure point. AWS CISO Amy Herzog’s current strategy focuses on reducing the manual guidance required from engineers to deliver actionable results, effectively shifting the security burden from human intuition to automated reasoning.

Proactive Hunting: Deconstructing Project Glasswing

Project Glasswing represents a shift toward systemic vulnerability hunting. Rather than relying on traditional static analysis or scheduled penetration tests, AWS is integrating Claude Mythos Preview—Anthropic’s most advanced model—directly into its continuous security review pipeline. The goal is to identify architectural weaknesses in mission-critical workloads before they are exposed to the public internet.

The technical efficacy of this approach is measured by the reduction in manual engineer intervention. According to internal testing, Claude Mythos Preview has proven more productive than previous iterations at surfacing security findings that were previously missed in well-tested environments. This suggests a move toward a self-healing code architecture where AI agents act as a persistent, automated red team.

“Project Glasswing embodies this approach by using AI to systematically hunt for vulnerabilities across critical infrastructure before adversaries find them.” — Amy Herzog, AWS CISO

As organizations scale their cloud footprints, the complexity of managing these permissions often leads to over-privileged accounts. To mitigate this, enterprise IT departments are increasingly relying on cybersecurity auditors and penetration testers to validate that their AI-driven defenses aren’t leaving blind spots in their specific implementation.

The IAM Bottleneck: Managing 1.2 Billion Calls Per Second

The sheer scale of AWS Identity and Access Management (IAM) is a study in distributed systems pressure. Handling 1.2 billion API calls per second requires more than just raw compute; it requires a logic engine capable of making instantaneous permit/deny decisions without introducing significant latency into the request lifecycle.

The latest update to IAM Access Analyzer introduces “internal access findings.” This tool leverages automated reasoning to analyze policy types, including service control policies and resource control policies. By mapping the relationship between roles, users, and resources like Amazon Relational Database Service (RDS) snapshots or DynamoDB tables, the tool identifies unintended access paths that could be exploited for lateral movement within a VPC.

For developers looking to audit their current access analyzer findings via the AWS CLI, the following command structure is used to list findings for a specific account:

aws access-analyzer list-findings --analyzer-arn arn:aws:access-analyzer:region:account-id:analyzer/analyzer-id

This level of granular control is essential for maintaining SOC 2 compliance and ensuring that containerization strategies don’t inadvertently expose sensitive data stores to the wider network. However, implementing these policies at scale often requires the expertise of managed service providers (MSPs) who specialize in cloud governance and IAM policy hardening.

Mitigating the S3 Blast Radius

The threat landscape for object storage remains aggressive. In 2025 alone, AWS blocked over 300 million attempts to maliciously encrypt customer files on Amazon S3. This volume of attacks highlights a persistent trend: adversaries are targeting the data layer directly, bypassing traditional network perimeters to execute ransomware-style encryption at the storage level.

To counter this, AWS is deploying a layered defense strategy that combines AI-powered log analysis with network-level protection. The introduction of the AWS Shield network security director (currently in preview) aims to simplify the management of distributed denial of service (DDoS) protections, reducing the time it takes for SecOps engineers to respond to emerging threats.

Security Capability Comparison: Traditional vs. AI-Driven

The architectural flow here is clear: by utilizing Anthropic’s models for safety research and foundation model development, AWS is creating a feedback loop where the AI is used to secure the very infrastructure that hosts the AI. This symbiotic relationship between the cloud provider and the model developer is designed to eliminate the latency between the emergence of a threat and the deployment of a defense.

the success of Project Glasswing and the IAM updates will be judged not by the number of threats blocked, but by the reduction in “time-to-remediation.” As the industry moves toward autonomous security agents, the role of the human engineer shifts from writing the rules to auditing the AI’s reasoning. For firms struggling to keep up with this transition, partnering with specialized software development agencies to integrate security-as-code into their CI/CD pipelines is no longer optional—it is a requirement for survival.