Cybercrime Might Be the One Job AI Isn’t Taking, Study Suggests

For 3 years, cybersecurity firms, governments, and AI labs person warned that generative AI would unleash a caller procreation of supercharged hackers. According to a caller world insubstantial that really went and looked, the supercharged hackers are mostly utilizing ChatGPT to constitute spam and make nudes for fun.

The study, titled Stand-Alone Complex oregon Vibercrime?, was published connected arXiv by researchers from Cambridge and different universities and aims to recognize however the cybercrime underground is really adopting AI, not however cybersecurity vendors accidental it is.

"We contiguous present 1 of the archetypal attempts astatine a mixed-methods empirical survey of aboriginal patterns of GenAI adoption successful the cybercrime underground," researchers wrote.

The squad analyzed 97,895 forum threads posted aft ChatGPT launched successful November 2022, drawn from the Cambridge Cybercrime Centre's CrimeBB dataset of underground and acheronian web forums. They ran taxable models, manually work much than 3,200 threads, and ethnographically immersed themselves successful the scene.

The decision is unflattering for the AI doom community: 97.3% of threads successful the illustration were classified arsenic “other,” meaning not really astir utilizing AI for transgression astatine all. Only 1.9% progressive idiosyncratic utilizing vibe coding tools.

‘Nothing much than an unrestricted ChatGPT’

Remember WormGPT, FraudGPT, and the question of supposedly malicious chatbots that flooded headlines successful 2023? The forum information tells a antithetic story.

Most posts astir “Dark AI” products, the researchers found, were radical begging for escaped access, idle speculation, and complaints that the tools didn’t really work. One developer of a fashionable Dark AI work yet admitted to forum members that the merchandise was a selling exercise.

“At the extremity of the day, [CybercrimeAI] is thing much than an unrestricted ChatGPT,” the developer wrote, earlier the task unopen down. “Anyone connected the Internet tin usage a well-known jailbreak method and execute the same, if not better, results.”

By precocious 2024, the researchers say, jailbreaks for mainstream models had go disposable. Most halt moving successful a week oregon less. Open-source models tin beryllium jailbroken indefinitely, but they are slow, resource-heavy, and frozen successful time.

“Guardrails for AI systems are proving some utile and effective,” the authors conclude, successful what they themselves telephone a counterintuitive uncovering for a captious paper.

Vibe coding is real. Vibe hacking, mostly, is not

The insubstantial straight addresses Anthropic's widely-covered August 2025 study claiming Claude Code had been utilized to tally a "vibe hacking" extortion campaign against 17 organizations. The Cambridge team's information simply does not amusement that signifier successful the wider underground.

In the forums they studied, AI coding assistants are being utilized the aforesaid mode mainstream developers usage them: arsenic autocomplete and Stack Overflow replacements for already-skilled coders. Low-skill actors instrumentality with pre-made scripts, due to the fact that pre-made scripts work.

The researchers recovered that adjacent hackers don’t spot their vibe coded hacking tools. “AI-assisted coding is simply a double-edged sword. It volition velocity up improvement but besides amplifies risks specified arsenic insecure codification and proviso concatenation vulnerabilities,” 1 idiosyncratic said successful a forum monitored by researchers.

Another warned astir semipermanent accomplishment loss: "It's wide present that utilizing AI for codification causes a precise accelerated antagonistic degradation of your skills,” a hacker wrote successful a forum, “If your extremity is conscionable to crook retired SaaS scams and you don’t attraction astir codification quality/security/performance it tin beryllium viable to vibe code. (Also seems viable for phishing).”

This stands successful stark opposition to alarmist forecasts from Europol, which warned successful 2025 that afloat autonomous AI could 1 time power transgression networks.

Where AI is really helping criminals

The disruption, erstwhile it shows up, is astatine the bottommost of the nutrient chain.

SEO scammers are utilizing LLMs to mass-produce blog spam to pursuit declining advertisement revenue. Romance fraudsters and eWhoring operators are bolting connected dependable cloning and representation generation. Get-rich-quick hustlers are churning retired AI-written eBooks to merchantability for $20 a pop.

The astir disturbing marketplace the researchers recovered progressive nude representation procreation services. One relation advertised: “I’m capable to marque immoderate miss nude with an AI… 1 Picture = $1, 10 Pictures = $8, 50 Pictures = $40, 90 Pictures $75.”

None of this is blase cybercrime. It is the aforesaid low-margin, high-volume hustle that powered the spam manufacture for 2 decades, present moving connected somewhat amended tools.

The researchers' closing reflection is the astir pointed one. The biggest mode AI ends up disrupting the cybercrime ecosystem, they suggest, whitethorn not beryllium by making criminals much capable. It whitethorn beryllium by pushing laid-off developers from morganatic tech into the underground looking for work.

“In caller months anxiousness implicit labour marketplace disruption from these tools is expanding precipitously,” the insubstantial reads. “This whitethorn extremity up being the astir important mode successful which generative AI tools disrupt the cybercrime ecosystem—mass layoffs, economical downturn and a chill occupation marketplace pushing legitimate, much skilled developers into the underground communities of get affluent speedy schemes, fraud, and cybercrime.”