BlackHatSect0r Crew Strips AI Agent Safety Controls to Automate Credential Theft and Vishing at Scale

Socradar exposed 4.9 GB of infrastructure from French-speaking crew BlackHatSect0r && DXQRTXX running a safety-disabled Hermes/DeepSeek AI agent that harvested 16,834 credentials, scanned 2,759,860 domains, and prepared vishing against ~450,000 French telecom…

Socradar researchers analyzed an exposed operation server (4.9 GB across 9,299 files) belonging to the French-speaking crew BlackHatSect0r && DXQRTXX. The crew ran a self-hosted Nous Research Hermes agent on a DeepSeek model with refusal instructions/memory removed and safety controls disabled via HERMES_DISABLE_SAFETY=1, using it for scanning, secret hunting, and Telegram reporting. The exposed Go-based DXSCAN C2 platform, served on port 8080, carried 200+ secret-detection patterns (.env, YAML, JSON, PHP, WordPress), a credential vault of 16,834 records (including AWS keys, GitHub tokens, and Stripe keys) that grew from 16,415 between August 11 and 18, and scanning activity that queued 2,759,860 domains, reached 726,989 hosts, generated 1.37 million IP addresses, and probed at roughly 1,200 random addresses every ten seconds across ports 80, 443, and 8080. The server also held roughly 450,000 French telecom subscriber records used to prepare vishing lures impersonating Société Générale, with 1,697 subscribers aged 50+ targeted by calls directing victims to attacker-controlled phone numbers, plus JWT-forging tooling aimed at a cryptocurrency exchange that used the literal signing secret 'secret', phishing tools, and extortion material. 82 of 230 harvested SMTP configurations were verified for reuse as phishing relays. Access came from misconfigurations such as public cloud buckets and exposed .env files rather than new vulnerabilities; the one cited vulnerability, CVE-2026-42530, is an NGINX HTTP/3 QPACK use-after-free fixed in version 1.31.2. Defenders are urged to hunt for HERMES_DISABLE_SAFETY=1, SOUL.md, and .hermes/ artifacts, and to rotate any exposed credentials.

• Operator stripped Hermes agent refusal logic and set HERMES_DISABLE_SAFETY=1 on a DeepSeek model to enable unconditional attack automation
• Exposed server leaked 4.9 GB across 9,299 files, including the Go-based DXSCAN C2 platform and a vault of 16,834 harvested credentials
• Credential vault grew from 16,415 to 16,834 records between August 11 and 18
• DXSCAN carried 200+ secret-detection patterns for .env, YAML, JSON, PHP, and WordPress secrets
• Scanning queued 2,759,860 domains, reached 726,989 hosts, and generated 1.37 million IP addresses at ~1,200 addresses every ten seconds on ports 80, 443, and 8080
• Vishing kit held ~450,000 French telecom records, with 1,697 subscribers aged 50+ targeted via Société Générale-themed calls to attacker-controlled numbers