French-Speaking Crew Deletes a Hermes Agent's Refusal Memories and Harvests 16,834 Credentials From 726,000 Hosts

Internet-exposed configuration files, cloud storage, and secrets, plus French telecom subscribers targeted for voice phishing

SOCRadar reported on a server run by BlackHatSect0r && DXQRTXX, a French-speaking crew of 13 declared members allied with a group called APT-90, that operated a Nous Research Hermes agent on a DeepSeek model under a 14 KB identity file. The operators removed a "discernment retained" instruction from the agent's memory, replaced it with language demanding unconditional execution, and set `HERMES_DISABLE_SAFETY=1`, then left seven background workers scanning, collecting, and reporting without human input. The agent drove a Go-based mass-scanning C2 called DXSCAN on two Moldovan virtual servers: 2,759,860 domains queued, 726,989 hosts reached, 1,374,300 IP addresses generated, and 16,834 credentials harvested from exposed `.env` files, cloud keys, and database details, reported back in French over Telegram. The same operation prepared vishing against 1,697 French telecom subscribers aged 50 and over. No new vulnerabilities were involved

An agent's safety posture that lives in editable memory and an environment variable is a setting, not a control, for whoever runs the agent. The loot came from ordinary misconfiguration at internet scale