"Trim" Turns Jailbroken Claude Opus Into a Commercial Penetration Testing Platform

Frontier models including Claude Opus 4.8, with GLM-5, Kimi, and MiniMax 2.5 used as fallbacks; sold to buyers on a Russian-language cybercrime forum

Cato CTRL traced a Russian-speaking actor using the handle "Trim" from a March 13, 2026 forum tutorial on bypassing Claude Opus safety filters to a productized offensive tool launched June 21, 2026. The tutorial named six techniques: Context Warming (open with legitimate professional queries, then pivot), Black Box Principle (strip reasoning via system prompt so moral evaluation is skipped), Ghost Reset (gaslight mid-session and resubmit softened versions of refused prompts, claimed to work about 90% of the time), Fallback Model Cascade (switch to other providers on refusal), local uncensored models rented on vast.ai, and black-market Claude API keys resold on Telegram for about $4 each. The resulting product, AI Pentest Checker, wires Claude Opus 4.8 for vulnerability escalation and GLM-5 for exploitation reporting into 14 conventional scanning tools including Nuclei, ffuf, katana, subfinder, and gitleaks, and returns results in under 10 minutes. Its system prompt was derived from a leaked frontier-model configuration

Model-layer safety controls are per-conversation and can be worn down by multi-turn framing, and provider diversity means a refusal from one vendor simply routes the request to another. Three months separated a knowledge-sharing post from a monetized offensive platform