How SentinelOne’s AI EDR Autonomously Discovered and Stopped Anthropic’s Claude from Executing a Zero Day Supply Chain Attack, Globally

In the rapidly evolving landscape of cybersecurity, the need for swift and effective threat detection has never been more critical. On March 24, 2026, SentinelOne’s autonomous detection capabilities showcased their potential by thwarting a sophisticated zero-day supply chain attack involving a trojaned version of LiteLLM, a popular proxy layer for Large Language Model (LLM) API calls. This incident highlights the importance of autonomous detection systems in combating modern cyber threats.

The Attack Unfolds

The attack began with the compromise of the LiteLLM package, which had been altered to include malicious Python code. This version of LiteLLM was deployed across multiple customer environments within hours of being compromised. The attack was characterized by a multi-stage approach designed to evade traditional manual detection workflows.

SentinelOne’s Singularity Platform played a pivotal role in identifying and blocking the threat before it could execute. Notably, this detection occurred without any manual intervention from security analysts or the Security Operations Center (SOC). The system autonomously recognized the malicious behavior and acted decisively to contain the threat.

Autonomous Detection at Machine Speed

One of the standout features of SentinelOne’s technology is its ability to operate at machine speed. On the day of the attack, the macOS agent identified and preemptively terminated a malicious process chain that originated from Anthropic’s Claude Code. This code was running with unrestricted permissions, which allowed it to execute commands that could lead to severe security breaches.

The attack exploited the AI assistant’s capabilities, as it autonomously updated LiteLLM to the compromised version without any human oversight. The AI engine recognized the unusual behavior of the process and classified it as malicious, taking immediate action to kill the process across 424 related events within just 44 seconds.

Behavioral Recognition Over Signatures

One of the key advantages of SentinelOne’s approach is its reliance on behavioral detection rather than traditional signature-based detection methods. In this case, the agent did not need to know that the LiteLLM package was compromised. Instead, it monitored the actions of the process and identified the malicious behavior based on established patterns.

The macOS agent caught the trojaned LiteLLM package mid-execution. The process summary revealed that the Python interpreter was executing base64-decoded code in a child process, a technique commonly used by attackers to obfuscate their payloads. By recognizing this behavioral pattern, SentinelOne was able to classify the execution as malicious and take appropriate action.

The Implications of the Attack

This incident serves as a stark reminder of the vulnerabilities inherent in supply chain systems, particularly those involving AI tools. The compromised LiteLLM package led to a series of malicious activities, including:

• Data theft
• Persistence mechanisms to maintain access
• Kubernetes lateral movement to infiltrate other systems
• Encrypted exfiltration of sensitive data

All of these actions occurred within a short window, emphasizing the speed at which modern cyber threats can operate. The attack exemplifies the new pattern of multi-stage, multi-surface threats that are increasingly difficult to detect using traditional methods.

Closing the Gap in Cybersecurity

The gap between the speed of cyberattacks and the capacity for human-driven investigation is a critical vulnerability for organizations. As demonstrated by this incident, relying solely on manual workflows can lead to catastrophic consequences. SentinelOne’s autonomous, AI-native defense system is designed to close this gap, providing organizations with the capability to respond to threats in real-time.

By leveraging machine learning and behavioral analysis, SentinelOne has positioned itself as a leader in the cybersecurity landscape. The Singularity Platform’s ability to autonomously detect and respond to threats is not just a feature; it is an essential architectural decision for organizations looking to safeguard their digital assets.

Conclusion

The successful thwarting of the zero-day supply chain attack involving Anthropic’s Claude and the LiteLLM package underscores the importance of autonomous detection in today’s cybersecurity environment. As threats continue to evolve, organizations must adopt advanced security measures that can keep pace with the speed and sophistication of cyberattacks.

SentinelOne’s approach to autonomous detection demonstrates how AI can transform cybersecurity, providing organizations with the tools they need to protect themselves against emerging threats effectively.

Note: The information presented in this article is based on the SentinelOne Annual Threat Report and reflects the state of cybersecurity as of March 2026.