AI-Assisted Cyberattacks Reshape Risk Assessment Frameworks, Study Reveals

The threat intelligence landscape is undergoing a seismic shift as new research from Anthropic indicates that traditional frameworks for assessing cyberattack risks are becoming obsolete. This change is not due to flaws in the frameworks themselves, but rather the transformative impact of artificial intelligence (AI) on the dynamics of attacker skill and capability. As AI-assisted cyberattacks become increasingly prevalent, the implications for cybersecurity are profound.

Shifting Paradigms in Cyber Threat Assessment

Anthropic’s Frontier Red Team has released findings based on an analysis of 832 accounts banned for malicious cyber activities between March 2025 and 2026. These cases were mapped against the MITRE ATT&CK framework, which is widely regarded as the industry standard for categorizing attacker tactics and techniques. The report highlights a troubling trend: the methodologies used to evaluate the sophistication of cyber actors are failing to keep pace with the evolving threat landscape.

A subset of these findings was included in Verizon’s 2026 Data Breach Investigations Report. The comprehensive analysis, which also featured an interactive visualization on Anthropic’s Red blog, presents three alarming conclusions that challenge the effectiveness of traditional risk-scoring models employed by security teams.

AI’s Expanding Role in Cyberattacks

The most prevalent AI-assisted activity identified in the dataset is the creation of malware. Of the 832 accounts examined, 560, or 67.3%, utilized AI for this purpose. However, a more significant trend emerged over the twelve-month study period: the application of AI shifted from initial-access techniques to more complex post-compromise activities. This shift indicates a growing trend toward sophisticated operations that occur after an attacker has already infiltrated a network.

AI-assisted account discovery, which involves identifying valid accounts within a compromised environment, increased by 8.9% during the study period. Conversely, AI-assisted phishing—a common initial-access technique—declined by 8.6%. Additionally, AI was employed in lateral movement, a technique used to navigate deeper into a compromised network to reach high-value targets, by 54 of the 832 actors, or 6.5%. These techniques, historically requiring advanced skills, are now becoming accessible to a broader range of attackers due to AI.

The risk-scoring data underscores this democratization of cyber capabilities. In the first half of the study, 33% of actors were classified as medium risk or higher. By the second half, this figure surged to 56%, marking a 1.7-fold increase in just six months.

The Ineffectiveness of Traditional Risk Signals

Historically, security teams have gauged an actor’s sophistication by counting the distinct techniques they employ and noting the tools or interfaces used. However, Anthropic’s data reveals that these signals have become disconnected from actual risk in an AI-driven context.

The least skilled actors in the dataset employed an average of 16 distinct techniques, while the most skilled used an average of 20. This minimal gap renders traditional assessments ineffective for operational triage. Furthermore, the type of interface used—whether Claude Code, an API, or a chat interface—showed no correlation with the level of risk.

What does differentiate higher-risk actors is their application of AI throughout the attack lifecycle. These operators focus their AI efforts on demanding techniques such as account discovery, lateral movement, and privilege escalation, rather than solely on initial-access tasks. However, this distinction is diminishing as more actors are reclassified as higher risk, with behaviors diffusing throughout the threat actor ecosystem.

Anthropic’s research indicates that the most dangerous actors construct frameworks around AI models that enable them to chain together various stages of a cyberattack with minimal human intervention. This capability, referred to as agentic attack orchestration, represents a new frontier in AI-enabled cyber threats and is not adequately captured by the current MITRE ATT&CK framework.

Gaps in the MITRE ATT&CK Framework

The limitations of the MITRE ATT&CK framework are starkly illustrated by a state-sponsored cyber espionage operation disrupted by Anthropic in November 2025. In this instance, a malicious actor manipulated Claude Code to infiltrate targets globally with minimal human oversight.

Mapping this operation against the MITRE ATT&CK framework revealed 30 techniques across 13 tactics—a profile comparable to many medium-risk actors in the dataset. However, this assessment significantly underrepresents the actual threat posed by the operation. Anthropic’s risk-scoring methodology assigned a maximum score of 100 to the same operation.

The discrepancy arises because MITRE ATT&CK was designed to document what attackers do, not how they orchestrate their actions. An AI agent capable of executing commands, exploiting vulnerabilities, stealing credentials, and making real-time tactical decisions across an entire attack chain—with human input required only at critical junctures—constitutes a fundamentally different threat than a human operator executing the same steps manually. Currently, there is no ATT&CK ID for agentic orchestration, nor is there a technique entry for the autonomous chaining of attack stages.

Anthropic is actively engaging with MITRE to explore how the ATT&CK framework can evolve to encompass these AI-enabled behaviors. The findings from this analysis are also informing the cyber safeguards integrated into Anthropic’s advanced models, including mechanisms for detecting and blocking malware development and mass data exfiltration activities documented in the dataset.

Urgent Need for New Detection Frameworks

The findings underscore that risk triage models based on technique counts, tool-type signals, or initial-access sophistication are systematically underclassifying AI-enabled actors. A threat actor employing 16 techniques with AI assistance may present the same operational risk as one using 25 techniques manually. Similarly, an attacker utilizing a free-tier chat interface may be executing the same agentic attack chain as one using a direct API connection.

The critical questions for detection and triage are now behavioral and architectural. Is the actor employing AI post-compromise rather than solely for initial access? Is there evidence of automated chaining between attack stages? Is human intervention being minimized in operationally demanding steps? These questions have yet to be integrated into standard detection frameworks, and addressing this gap is an urgent priority for the cybersecurity industry.

For further insights into the evolving landscape of AI-assisted cyber threats, visit thecyberexpress.com.

Keep reading for the latest cybersecurity developments, threat intelligence and breaking updates from across the Middle East.