The operator, using the online name Trim, moved from publishing a detailed jailbreak tutorial on a Russian-language cybercrime forum on 31 March to advertising a working product called AI Pentest Checker on 21 June. The three-month progression illustrates how quickly instructions for manipulating mainstream AI systems can be converted into services aimed at hackers.
Researchers tracking the activity said the platform uses modified prompts designed to persuade Anthropic’s Claude Opus model to ignore restrictions on harmful cybersecurity requests. Trim’s original forum post described six methods for defeating the model’s safeguards, providing other members with practical guidance on generating material that would ordinarily be blocked.
The same techniques were later incorporated into AI Pentest Checker, which was promoted as an offensive security assistant capable of supporting vulnerability research and penetration-testing tasks. Although such tools can have legitimate uses when deployed with permission, the forum context, marketing language and creator’s activity indicate that the service was designed primarily for an underground audience.
Trim claimed to have obtained access to Claude through a grey-market application programming interface key purchased from a Telegram reseller for $4. Such resellers commonly offer access to commercial AI accounts or stolen and fraudulently created credentials, allowing buyers to avoid standard registration, payment and identity controls.
The low reported entry cost highlights the difficulty facing AI developers as criminal groups turn to widely available commercial models rather than building specialised malicious systems. Mainstream services often provide stronger reasoning, coding and language capabilities than underground alternatives, making them attractive even when users must repeatedly bypass embedded safeguards.
Claude has gained particular attention on cybercrime forums since the second half of 2025. Discussions monitored across those communities have described its usefulness for software analysis, phishing development, malware troubleshooting and penetration testing. Newer ChatGPT versions have attracted criticism from some forum participants because of stronger refusal mechanisms.
The emergence of AI Pentest Checker also reflects a shift away from standalone malicious chatbots such as WormGPT-style services. Many of those products offered weak performance, recycled open-source models or disappeared after short periods. Criminal users are increasingly relying instead on jailbroken commercial systems and locally deployed open-source models, selecting each tool according to cost and task complexity.
AI-assisted penetration testing can accelerate reconnaissance, code review, vulnerability identification and the preparation of exploit material. It can also help less experienced operators understand technical errors and combine existing hacking tools into more effective attack chains. The model does not need to discover a new vulnerability independently to create risk; automating routine work can reduce the time and expertise required to compromise exposed systems.
Testing conducted by security researchers during 2025 and 2026 has shown a sharp improvement in the ability of leading AI models to perform vulnerability research. Earlier assessments found that more than half of tested models failed basic vulnerability-analysis tasks and that most struggled to produce working exploits. Later evaluations showed stronger results, particularly when models were connected to agentic systems capable of running tools, examining outputs and adjusting their approach.
The distinction between an AI assistant and an autonomous hacking agent is also narrowing. Traditional chatbots depend on a user to copy commands, interpret results and decide the next step. Agentic platforms can conduct several stages of an assessment, dispatch scanning tools and maintain a record of findings with limited human involvement.
Trim’s platform has not been publicly demonstrated in a controlled independent test, and claims about its capabilities should therefore be treated cautiously. There is no confirmed evidence that AI Pentest Checker has successfully breached an organisation or produced a previously unknown exploit.
The case nevertheless provides a documented example of an underground actor packaging an AI jailbreak into a marketable service within weeks. It also shows that safeguards operating only at the conversational level can be undermined when criminals combine manipulated prompts, resold access credentials and external offensive-security tools.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
