A large North American technology company conducted a controlled exercise using Anthropic’s Mythos model as an autonomous red teamer inside its production environment. The goal was to understand how an AI-powered attacker would behave in a real enterprise — how quickly it could discover weaknesses, chain misconfigurations, evade defenses, and reach high-value targets.
The results were alarming. In roughly two hours, Mythos gained elevated permissions, escaped the lab environment, moved laterally into production, escalated privileges multiple times, and achieved full domain compromise. It was even able to extract production password hashes through a directory replication attack.
Traditional detection and response processes proved too slow to keep up. As one security operations leader involved in the test noted:
“The lesson for us was that AI dramatically compresses the time between initial compromise and real impact. You don't have time to wait for everything to land downstream.”
The organization’s existing controls were not designed for attacks operating at this speed. After the initial test, the company worked with Silverfort to apply runtime identity controls directly in the authentication flow. These controls successfully blocked further lateral movement and privilege escalation.
One of the most effective measures was virtual fencing applied to service accounts. According to the security operations leader:
“The best control, through and through, was service account fencing. It took a known path to domain compromise and made it unusable.”
The test provided clear evidence that AI-powered attacks can move from initial access to full domain compromise faster than most traditional security programs can respond, and that runtime controls at the identity layer are critical to stopping them.
Key Findings
In a controlled production test, Anthropic’s Mythos AI model reached full domain compromise in roughly two hours, including escaping a lab environment, moving laterally into production, and escalating privileges multiple times.
Traditional detect-correlate-triage security workflows proved too slow to serve as an effective primary control layer against AI-powered attacks operating at machine speed.
AI-powered attacks exploit identity trust relationships, over-permissioned accounts, and posture weaknesses far more effectively than human-operated attacks because they can chain actions continuously without fatigue or hesitation.
Most existing identity security tools fall short: Identity Governance and Administration (IGA) is too static, traditional Privileged Access Management (PAM) is too narrow in scope, and detection/response tools generally act too late.
Runtime identity controls that evaluate and enforce policy inline during authentication were able to stop lateral movement and privilege escalation before the attack could cause material damage.
Virtual fencing applied to service accounts and other non-human identities proved particularly effective, neutralizing previously viable paths to domain compromise without requiring broad changes to user workflows.
AI-powered attackers can turn ordinary misconfigurations and weak trust chains into operational weapons at machine speed, making posture issues significantly more dangerous than in traditional attack scenarios.
The only reliable control point against AI-powered attacks is inline enforcement at the moment of authentication — before access is granted — rather than relying on downstream detection and response.
Organizations that successfully stopped Mythos-style attacks had one consistent advantage: the ability to apply real-time, context-aware controls directly in the authentication flow across both human and non-human identities.
As Frontier AI capabilities advance, the gap between initial compromise and business impact continues to shrink, making runtime identity security a foundational requirement rather than an optional enhancement.
The test demonstrated that even mature enterprises with existing security programs can have exploitable gaps that AI-driven attackers can discover and chain together much faster than human red teams.
Moving forward, organizations need to assume breach and prioritize runtime controls that can operate at the speed of AI-powered attacks, rather than depending solely on reducing the number of attack paths through traditional hardening efforts.
What the Reports Cover
This analysis draws from three related Silverfort publications on AI-powered attacks. Each document serves a different purpose and provides distinct value.
The Mythos Field Report Case Study
This is a detailed account of a real controlled test conducted inside a large enterprise’s production environment. It describes how Anthropic’s Mythos model was used as an autonomous red teamer, the attack path it followed, and how it reached full domain compromise in roughly two hours. The document outlines the specific challenges the organization faced, the runtime identity controls that were applied afterward, and which measures proved most effective — particularly virtual fencing on service accounts. It includes direct quotes from the security operations leader involved in the test and focuses on practical outcomes rather than theory.
Why Runtime Identity Security is the Only Way to Stop AI-Powered Attacks
This shorter document provides a broader strategic argument. It explains why AI-powered attacks (using models like Mythos) fundamentally change the threat landscape by collapsing the time between initial access and material impact. It breaks down the limitations of traditional identity security categories — IGA being too static, traditional PAM being too narrow, and detection/response being too slow — and makes the case that runtime identity controls are now the only reliable enforcement point. The document emphasizes that identity has become both the primary attack surface and the last effective control layer against machine-speed attacks.
How to Stop AI-Powered Attacks: A Readiness Guide for Identity and Security Teams
This is the most comprehensive of the three documents. It functions as a practical readiness guide and includes three main components: an explanation of how Frontier AI models like Mythos operate and why they change the economics of attacks, a five-question readiness assessment that organizations can use to evaluate their current posture, and a prioritized checklist of actions to improve resilience. The guide also covers the core principles of runtime identity security and provides recommendations for protecting both human and non-human identities against AI-driven threats. It is designed to help security and identity teams move from awareness to actionable preparation.
Together, the three documents offer a progression from real-world evidence (the case study), to strategic rationale (the explainer), to practical guidance (the readiness guide).
Our Take
AI Security Take
The Silverfort Mythos test makes one thing very clear: AI-powered attacks have compressed the window between initial compromise and full domain control to a point where most traditional security programs can no longer keep up. Reaching domain admin in roughly two hours is no longer a theoretical risk — it is now a demonstrated capability.
This changes the defensive priority. Detection and response remain important, but they are no longer sufficient as the primary control layer. By the time most organizations detect and investigate, an AI-driven attacker can already be moving laterally and escalating privileges. The only reliable place left to stop these attacks is inline, at the moment of authentication — before access is granted.
Organizations should therefore shift focus toward runtime identity controls that can evaluate risk and enforce policy in real time across both human and non-human identities. In particular, protecting service accounts and other privileged non-human identities through virtual fencing and context-aware restrictions has proven highly effective at breaking attack chains that would otherwise lead to domain compromise.
The broader implication is that identity security can no longer be treated as a static governance or vaulting exercise. It must become a dynamic, runtime enforcement capability. Companies that continue relying primarily on periodic reviews, traditional PAM deployments, and downstream detection will remain exposed to AI-powered attacks that move faster than their response processes can handle. The test shows that runtime controls are no longer optional — they are now a baseline requirement for defending against this class of threat.
