Research indicates a transition from AI-assisted to AI-embedded malware through the integration of 8B-parameter Dolphin-family Small Language Models (SLMs) into Remote Access Trojans (RATs). These agentic RATs utilize quantized local inference engines, such as LM Studio, to execute an autonomous "Observe-Decide-Act" (ODA) loop on compromised commodity hardware. By performing reasoning locally, the malware reduces dependency on Command & Control (C2) communication and cloud APIs, effectively minimizing network-based telemetry and bypassing traditional EDR/NDR detection. While current operational reliability is constrained by model hallucinations (~10.9% success rate), the architectural feasibility of achieving autonomous root-shell access represents a Tier 3 sophistication level in modern offensive AI.

• Threat Model: AI-Embedded Autonomy

  • Shift from "AI-assisted" (human-led code generation) to "AI-embedded" (malware-led execution).
  • Integration of 8B-parameter Dolphin-family SLMs via quantization for local endpoint execution.
  • Reduction of network-based indicators by eliminating frequent cloud API calls and C2 polling.

• Attack Mechanics: The ODA Loop

  • Implementation of a recursive "Observe-Decide-Act" (ODA) cycle to interpret host data and select tools.
  • Utilization of local inference engines like LM Studio to host the SLM "brain" on the compromised host.
  • Capability to perform autonomous reconnaissance, analyze command output, and self-correct without external guidance.

• Operational Performance & Reliability

  • Controlled testing demonstrates a 10.9% success rate on strict attack checklists using 8B models.
  • Validates the technical feasibility of achieving autonomous root-shell access on target environments like Metasploitable2.
  • Currently limited by high computational overhead and significant model hallucination rates.

• Defensive Implications & EDR Evasion

  • Agentic behavior generates non-linear command patterns that evade signature-based and heuristic EDRs.
  • Localized decision-making minimizes the "heartbeat" traffic used by traditional detection to identify RAT infections.
  • Requires a defensive shift toward behavioral monitoring of recursive, autonomous agentic cycles.

• Strategic Outlook

  • Trend suggests a convergence of standard RAT frameworks with SLM-based autonomous controllers.
  • Transition from conceptual research to practical application observed in the 2025-2026 threat window.
  • Moving toward self-contained, adaptive malware that reasons through exploitation in real-time.