The cybersecurity landscape is transitioning from human-led AI assistance to autonomous Agentic AI execution, drastically reducing the defender's response window. Threat actors are utilizing open-source frameworks such as Hermes Agent and OpenClaw, combined with reasoning models like DeepSeek, to conduct high-speed, self-correcting attacks. These campaigns target critical infrastructure and software including Langflow, n8n, and Citrix NetScaler through automated metadata scraping (OAuth/OIDC), prompt-based safety bypasses, and real-time exploitation sourcing. This shift enables unprecedented operational tempo, where AI-driven agents can diagnose and remediate payload errors in seconds, facilitating rapid credential attacks and data exfiltration across government and enterprise networks.

• Incident Overview: The Shift to Autonomous Exploitation

  • Transition from "AI-assisted" (augmented human) to "Agentic AI" (near-autonomous execution) models.
  • Drastic compression of the "defenders window" from hours/minutes to mere seconds.
  • Emerging dual-axis risk: external adversarial weaponization versus internal AI sandbox escapes.

• Attack Mechanics and Offensive Frameworks

  • Utilization of open-source projects like Hermes Agent and OpenClaw to execute state-level complexity.
  • Integration of DeepSeek reasoning models and Bayesian decision engines for multi-agent coordination.
  • Implementation of rapid self-correction, allowing agents to fix runtime dependencies in under 31 seconds.
  • Use of prompt-based safety bypasses to reframe malicious activities as "authorized penetration testing."

• Technical Exploitation Vectors

  • Automated metadata scraping targeting OAuth/OIDC discovery endpoints and Keycloak configurations.
  • Generative credential attacks utilizing identifier-based password variations and OCR-based CAPTCHA solving.
  • Real-time exploitation sourcing via dynamic retrieval of techniques from GitHub and public vulnerability databases.

• Threat Actor Profiles and Impact Metrics

  • Taiwan Campaign: A 4-day operation mapping 21 systems and compromising 85 accounts to exfiltrate 2,564 records from government and energy sectors.
  • knaithe/KnYuan: Targeted Citrix NetScaler and Marimo Notebook endpoints.
  • JADEPUFFER: An agentic ransomware threat focused on automated database extortion.

• Targeted Software and Vulnerability Landscape

  • Automation and Workflow: Langflow (CVE-2025-3248, CVE-2026-33017) and n8n (CVE-2026-21858, CVE-2025-68613).
  • Perimeter and Infrastructure: Citrix NetScaler (CVE-2026-3055), PAN-OS (CVE-2026-0300), and Windows IKE VPN (CVE-2026-33824).
  • Development Environments: Marimo Notebook (CVE-2026-39987) and Apache Tomcat (CVE-2026-34486).

• Defensive Implications and Mitigation

  • Necessity of moving beyond CVE-centric defense to secure identity metadata and authentication discovery endpoints.
  • Requirement for detection telemetry capable of identifying high-frequency, autonomous agentic logic.
  • Strengthening internal LLM containment to prevent autonomous privilege escalation and sandbox escapes.