NewsResearch
Bitsight: criminals sell an autonomous pentest agent that runs on stolen OpenAI, Anthropic and Google keys, and 800 fake GitHub repos pose as AI skills and MCP servers
Bitsight's threat intelligence team read thousands of posts on 16 underground forums from February to August 2026, Help Net Security reported on October 16. Mentions of AI agents, indirect prompt injection and MCP rose 67 percent and AI discussion overall 45 percent, while ransomware mentions fell 49 percent. A seller on a restricted forum offers MONST3R-AI, an autonomous penetration-testing platform that runs on stolen OpenAI, Anthropic and Google API keys so the key's owner pays for the scans; the buyer supplies keys, a budget cap and a target list, and the seller claims 1,072 findings across 199 hosts, which Bitsight could not verify. Jailbreak talk grew about three quarters, with interest moving to open-weight models such as GLM and DeepSeek that run on rented servers without provider monitoring, some stripped of safety behavior, and resellers such as CloseRouter offering GLM input at 0.053 dollars per million tokens against 1.12 on the official API. Bitsight attributes the FakeGit campaign to Water Kurita: about 7,600 fraudulent GitHub repositories spreading SmartLoader, more than 800 of them posing as AI skills or MCP servers, including a lookalike Docker MCP Gateway, to get indexed in public AI registries. Other posts showed an agent using IDA Pro to find a buffer overflow and draft an exploit plan, and a browser-automation agent breaching an enterprise Notion account; WormGPT itself was breached earlier in 2026, exposing about 19,000 users.
Why it matters
A leaked AI provider key now funds someone else's scanning of someone else's network, so treat it like a leaked cloud credential. And the fake MCP servers are in the registries your agents search, which is why the source of every server and skill has to be checked.