AI in cybercrime not turning hackers into superhackers
The study analyzed 97,895 forum threads posted after ChatGPT launched in November 2022; 97.3% of these threads were classified as “other,” meaning they were not actually about using AI for crime at all. This headline finding — that AI in cybercrime not turning hackers into superhackers — sets the scope for the article and frames the subsequent analysis of forum discussions following ChatGPT’s launch.
The Cambridge Cybercrime Centre at the University of Cambridge conducted an extensive study analyzing 97,895 forum threads to understand the role of AI in cybercrime following the launch of ChatGPT in November 2022. As part of their methodology, researchers manually examined 3,200 of these threads to ensure accurate classification. A key finding from the study revealed that 97.3% of threads fell into the “other” category, indicating they were unrelated to AI usage for criminal activities.
Furthermore, only 1.9% of the threads discussed the use of vibe coding tools, suggesting limited impact in the realm of cybercriminal activity. The study observed that AI coding assistants primarily serve as autocomplete tools and replacements for Stack Overflow among skilled coders, whereas low-skill individuals tend to resort to using pre-made scripts. This highlights a distinct difference in AI usage between experienced coders and novices within these forums.
By late 2024, the study reports that jailbreaks for mainstream AI models had become disposable. The researchers also report that open-source models could be jailbroken indefinitely, but these models were slow to run, required substantial computing resources, and were effectively frozen in time. These technical characteristics are presented in the study as concrete constraints on how jailbroken models functioned within the forums examined. The study documents the coexistence of ephemeral mainstream jailbreaks and persistent but resource-heavy open-source alternatives.
The Cambridge team’s dataset does not show a pattern of AI-driven extortion campaigns in the forums analyzed. The study also documents how AI coding assistants functioned in practice in those forums: as autocomplete and Stack Overflow replacements for already-skilled coders, while low-skill actors relied on pre-made scripts. The authors report these observed patterns alongside the technical limitations of jailbroken models, reporting that jailbreaking techniques existed in the ecosystem but came with practical constraints that affected their use in the sampled cybercrime discussions.
The Cambridge team’s dataset does not show a pattern of AI-driven extortion campaigns as claimed by Anthropic’s August 2025 report. This absence of evidence is reported within the study’s analysis of the forums sampled. The paper explicitly states that the dataset examined did not reveal the extortion campaign patterns described in the Anthropic report.
The authors conclude, ‘guardrails for AI systems are proving both useful and effective.’
They present this conclusion as part of their overall assessment of AI’s role in the forum discussions. The study frames the effectiveness of guardrails as a factor in limiting demonstrable AI-enabled criminal campaigns in the examined dataset.
The Cambridge Cybercrime Centre’s study adopts a cautious and analytical tone in presenting its findings on AI in cybercrime. The paper frames its analysis as debunking hype around AI-enabled cybercrime and bases conclusions on the forum data it examined. The study reports that AI in cybercrime not turning hackers into superhackers. The authors present their data as indicating that AI tools did not significantly enhance hackers’ skills or capabilities beyond existing levels.
