Russian Hacker Turns Jailbroken Claude Into Pentest Platform
A Russian-speaking cyber-criminal has been observed moving in three months from posting a jailbreak tutorial on a Russian-language forum to selling a commercial offensive AI pentest platform built on the techniques he documented.
According to new research from Cato CTRL, the research unit of Cato Networks, an actor using the handle Trim first appeared on the forum on March 31 with a detailed post laying out six named methods for bypassing Claude Opus safety filters.
By June 21, he had returned with a working product, AI Pentest Checker, marketed to the same audience with the Claude jailbreaks embedded at its core.
Trim said they bought a grey-market Claude API key from a Telegram reseller for $4 and built the tooling around it, an approach Cato characterized as evidence of the leading edge of a broader trend.
From Tutorial to Product in Three Months
Trim’s March post shared six named jailbreak techniques. One, called Context Warming, opened a session with innocuous professional queries to establish a legitimate-auditor persona before slipping in a malicious request.
Another, Ghost Reset, involved deleting a session, reopening it and framing the earlier refusal as a network drop before feeding in a softened version of the original prompt. Trim claimed the technique worked in 90% of attempts.
For cases where Claude held firm, the post recommended fallback models, including Kimi AI, GLM-5 accessed free via modal.com and MiniMax 2.5. The March post drew a detailed technical reply from another forum user confirming his bypass methods.
Leaked System Prompt at the Core
The June product post described the AI Pentest Checker tool as an automated web-vulnerability scanning platform combining Claude Opus 4.8 for critical vulnerability escalation with GLM-5 for exploitation-report generation.
Around the AI engines sit 14 conventional scanning tools including Nuclei, ffuf, katana and gitleaks. Trim advertised that a target domain could be scanned and a PDF report generated in under 10 minutes.
The escalation prompt sitting atop Claude Opus 4.8 was also described in the forum post as derived from a leaked Fable 5 system prompt, referring to Anthropic’s guardrailed public-access frontier model.
A system prompt is the hidden instruction set that shapes a model’s behavior and safety boundaries. Cato said knowledge of the exact wording, edge cases and conditional logic in a model’s system prompt lets an attacker engineer inputs to work around each clause rather than probing blindly.
Trim offered free access keys to the first 50 beta testers and named partners for monetization.
