AI-enabled cyber threats analysis by Anthropic's Frontier Red Team
Anthropic's Frontier Red Team analyzed a year's worth of AI-enabled cyber threats, mapping attacker techniques to the MITRE ATT&CK framework. The analysis reveals that malicious actors are increasingly using AI for complex, autonomous operations, challenging traditional risk assessment methods. These findings are informing the development of safeguards for Anthropic's models and discussions with MITRE to evolve the ATT&CK framework.
- →AI enhances cyber attacker capabilities and autonomy
- →Frontier Red Team analyzes AI-enabled cyber threats
- →MITRE ATT&CK framework may not fully capture AI-enabled threats
- →AI adoption in cyberattacks shifts towards post-compromise activities
- →Traditional risk assessment methods for cyber attackers are becoming less effective
Features (1) ›
- AI enhances cyber attacker capabilities and autonomy
Analysis shows malicious actors are using AI to become more dangerous, particularly in complex stages of cyber operations, leading to more autonomous attacks. This shift challenges traditional methods of differentiating high- from low-risk actors.
Enhancements (1) ›
- MITRE ATT&CK framework may not fully capture AI-enabled threats
The study found that the MITRE ATT&CK framework does not completely encompass the tools and activities that make AI-enabled attackers particularly dangerous. Behaviors like autonomous orchestration of attack stages are not yet fully represented.
Notes (4) ›
- Frontier Red Team analyzes AI-enabled cyber threats
Anthropic's Frontier Red Team investigated 832 accounts banned for malicious cyber activity between March 2025 and March 2026 to understand how AI is transforming cyberattacks. The analysis maps attacker techniques to the MITRE ATT&CK framework and highlights key findings regarding the evolving threat landscape.
- AI adoption in cyberattacks shifts towards post-compromise activities
AI-enabled activities in cyberattacks are increasingly focused on post-compromise actions like account discovery and lateral movement, rather than initial system access. This suggests attackers are applying AI deeper into the attack lifecycle, making sophisticated techniques accessible to less skilled actors.
- Traditional risk assessment methods for cyber attackers are becoming less effective
Traditional risk assessment based on the number of techniques employed or platform used no longer accurately reflects an attacker's threat level due to AI's ability to perform complex tasks. A more durable differentiator is the architectural scaffolding that enables AI models to chain attack stages with minimal human input.
- Findings inform AI model safeguards and framework evolution
The analysis has led to the deployment of cyber safeguards on Anthropic's models to detect and block malicious activities. Discussions are underway with MITRE to evolve the ATT&CK framework to include observed AI-enabled behaviors.
https://www.anthropic.com/news/AI-enabled-cyber-threats-mitre-attack