Anthropic Data Shows Top-Risk AI Attack Looked Ordinary by Technique Count

TechRadar reported that security operations have spent two decades getting good at recognizing evidence an attacker already used. Malware hashes, malicious IP addresses and suspicious domains gave teams a practical way to identify known threats and block repeat attacks. Once an attacker exposed part of their infrastructure, thousands of organizations could benefit. David Bianco's 2013 Pyramid of Pain outlined the weakness: the indicators defenders find easiest to consume are also the cheapest for attackers to discard. A hash can change when a file is modified, while an IP address can simply be replaced. The techniques and procedures an attacker relies on to reach an objective are far harder to change.

Generative AI has made indicator-based detection more vulnerable by driving down the cost of variation, according to the report. Code-generation tools can produce new malware variants quickly, phishing content can be created at scale, and offensive tooling can be assembled with far less specialist effort. Defenders increasingly work with indicators whose useful life may be shorter than the process required to identify, publish and act on them. The IOC feed remains valuable for blocking known-bad in volume, enriching investigations and hunting retrospectively once a campaign becomes public, but it can no longer carry the weight of the detection program, TechRadar reported.

Anthropic's study examined 832 accounts banned for malicious cyber activity between March 2025 and March 2026. It mapped 13,873 observed actions across 482 techniques and all 14 ATT&CK tactics. The accounts represent a subset of total bans with enough detail for thorough assessment. The study found little correlation between an actor's skill and the number of techniques they used: the least capable actors in the dataset averaged around 16 distinct techniques, and the most capable around 20. The platform an attacker worked through made no difference either.

That pattern tracks with how enterprise environments behave, according to the report. Account discovery appears in attacks, and also in legitimate administration. Remote services enable lateral movement and routine infrastructure management alike. Almost every individual technique requires interpretation before it supports a conclusion. The information appears when activities connect. Account discovery, followed by credential access, followed by movement into another system and data staging within a compressed window, describes something that the same activities spread across several days of routine work do not. Sequence becomes part of the detection logic, along with the identity involved, the systems touched and what happened immediately afterwards.

AI adds another consideration: the speed at which an attack unfolds. Automation can compress the time between actions, changing how otherwise familiar activity should be interpreted. Human-led attacks have always used some degree of automation, but an operator still has to interpret results and decide what to do next. As AI systems take on more of that decision-making, the time between stages of an attack can shrink considerably, making the speed and sequence of activity useful information in their own right.

Anthropic's data shows where this pressure is being applied, TechRadar reported. Across the study period, AI use shifted away from gaining access and toward what happens after. Use of AI tools for account discovery inside compromised environments rose 8.9 percent, while AI-assisted phishing fell 8.6 percent. Post-compromise techniques that once demanded real expertise are being performed on behalf of less capable actors. The proportion of actors Anthropic classified as medium risk or higher rose from 33 percent in the first half of the study to 56 percent in the second, roughly a 1.7-fold increase in twelve months. Speed alone proves nothing, according to the report.

Editor's Summary

Anthropic data examined by TechRadar shows that a state-sponsored campaign scored maximum risk despite using only 30 techniques, indicating that defenders need to focus on how techniques are combined and sequenced rather than on technique counts alone. The study also found AI use shifting toward post-compromise activity and a rising share of medium- or higher-risk actors. The IOC feed remains useful for blocking known threats, but the report says it cannot carry the weight of a detection program.