For years, cybersecurity teams have relied on established methods to determine how dangerous a threat actor might be. Analysts typically examine the techniques an attacker uses, the tools involved, and the complexity of an operation to estimate the level of risk. New research from Anthropic, however, recommends that artificial intelligence is beginning to disrupt those assumptions.
The company's Frontier Red Team recently analyzed 832 user accounts that were removed from Anthropic's platforms for engaging in malicious cyber activity between March 2025 and March 2026. Researchers compared the observed behavior against the MITRE ATT&CK framework, a widely used industry resource that categorizes adversary tactics and techniques. Portions of the findings were also referenced in Verizon's 2026 Data Breach Investigations Report.
It's a signal to keep up with how cybercriminals are using AI. Rather than limiting AI to basic tasks, attackers are increasingly applying it to activities that take place after gaining access to a target environment. This trend suggests that AI is becoming part of deeper operational stages of cyber intrusions, including tasks that traditionally required stronger technical expertise.
Among all observed cases, malware development was the most common use of AI. Researchers found that 560 of the 832 analyzed accounts, representing more than two-thirds of the dataset, used AI-assisted tools to help create or modify malicious software. While this finding was expected, the more notable change appeared elsewhere.
Throughout the study period, researchers recorded a movement away from AI-assisted initial access activities and toward post-compromise operations. One example was account discovery, a process attackers use to identify valid user accounts within a breached network. AI-assisted account discovery increased by 8.9% during the reporting period. By contrast, AI-supported phishing activity declined by 8.6%.
The data also showed growing use of AI during lateral movement operations. Lateral movement refers to the actions attackers take after entering a network to expand their access and reach more valuable systems, users, or data repositories. According to the report, 54 of the 832 observed actors used AI assistance during this stage of an intrusion.
Historically, activities such as account discovery, privilege escalation, and lateral movement have been associated with more experienced operators because they require a stronger understanding of network environments and attack workflows. Researchers argue that AI is reducing those technical barriers, allowing a broader range of actors to perform tasks that were previously more difficult to execute effectively.
This change became visible in the study's risk assessment data. During the first half of the observation period, approximately 33% of threat actors were categorized as medium-risk or higher. During the second half, that proportion rose to 56%. Researchers described this increase as evidence that AI is helping a larger segment of the threat landscape carry out more advanced cyber activity.
The findings also raise questions about how the industry evaluates attacker sophistication. Security teams have long treated the number of techniques used during an attack as an indicator of capability. Anthropic's analysis suggests that this relationship is becoming less reliable in AI-assisted environments.
Researchers found only a small difference between lower-risk and higher-risk actors when measuring the number of techniques used. Less sophisticated actors employed an average of 16 techniques, while the most capable actors averaged 20. The narrow gap indicates that technique counts alone may no longer provide a meaningful way to prioritize threats.
The same pattern appeared when researchers examined how attackers interacted with AI systems. Whether actors used Claude Code, direct API access, or standard chat interfaces showed little connection to their assessed risk level. Simply identifying which AI tool was used did not provide a clear indication of the threat posed by an actor.
Instead, researchers found that the location of AI usage within the attack lifecycle was a stronger indicator of risk. Higher-risk operators tended to apply AI to technically demanding stages of an intrusion, including internal reconnaissance, privilege escalation, and lateral movement. These activities often have a direct impact on how effectively an attacker can establish control over a compromised environment.
Even that distinction may not remain useful indefinitely. Researchers observed that these more advanced use cases are gradually spreading throughout the broader threat ecosystem. As AI tools become more accessible and capable, activities once associated with a smaller group of highly skilled operators may become increasingly common.
Anthropic identified another characteristic that separated the most dangerous actors from the rest. Rather than using AI for isolated tasks, some operators built systems around AI models that connected multiple attack stages together. This allowed AI to support planning, execution, and decision-making across larger portions of an operation with limited human involvement.
Researchers describe this capability as agentic attack orchestration. In practical terms, it refers to AI systems that can assist with coordinating different phases of an intrusion, helping move an attack from one stage to another without requiring constant manual direction from an operator.
According to the report, this rising behavior exposes a limitation in existing cybersecurity frameworks. MITRE ATT&CK was designed to document attacker actions and techniques. It was not built to measure the degree of autonomy involved when AI systems help coordinate those actions.
Anthropic underlined this challenge using a cyber-espionage campaign it disrupted in November 2025. The operation involved attempts to use Claude Code in support of intrusion activity targeting organizations in multiple regions with relatively little direct human intervention.
When researchers mapped the operation to MITRE ATT&CK, it generated a profile containing 30 techniques across 13 tactics. On paper, that profile appeared comparable to many medium-risk actors included in the study. However, Anthropic's internal evaluation system assigned the operation the maximum possible risk score of 100.
Researchers argue that the discrepancy exists because current frameworks focus on what actions occur during an attack rather than how those actions are coordinated. An AI-assisted system capable of executing commands, identifying vulnerabilities, collecting credentials, and adapting to changing conditions throughout an intrusion presents a different operational challenge than a human manually performing each step.
The report notes that there are currently no ATT&CK categories specifically designed to capture autonomous orchestration, automated chaining of attack stages, or the reduction of human decision-making throughout an attack lifecycle.
Anthropic says it is actively discussing potential framework updates with MITRE to better account for AI-enabled attack behaviors. The company has also used insights from the research to strengthen safeguards within its own models, including controls intended to detect and prevent misuse involving malware development and large-scale data theft attempts.
For defenders, the findings suggest that traditional indicators may no longer provide a complete picture of cyber risk. A threat actor using AI to automate portions of an attack may achieve outcomes similar to those of a more experienced operator performing the same tasks manually. Likewise, an individual using a basic chat interface could potentially conduct operations that resemble those performed through more advanced integrations.
