Systemic Coverage Gap Between Agentic LLM Attack and Defense Research (Literature Review)

Technical Analysis

The source is a PRISMA-guided systematic review categorizing agentic LLM vulnerabilities into a four-layer taxonomy: perception, brain, action, and interaction. It does not describe a specific exploit, payload, vulnerable framework version, or proof-of-concept; instead it aggregates findings across prior literature and highlights that action-layer vulnerabilities (tool misuse, code injection, sandbox escape) are underrepresented in research relative to their real-world risk. The paper's core observation is architectural: weak isolation between layers allows a compromised reasoning step (e.g., from prompt injection) to propagate into tool invocation, code execution, or persistent memory, escalating a perception-layer flaw into action-layer impact. No new attack technique, CVE, or affected product is disclosed; this is analytical/survey content useful for prioritizing defense research rather than an operational threat.

Detection Signatures

• N/A - this is a literature review with no specific IOCs, payloads, or exploit signatures to detect.

Remediation Steps

1. 1

   Prioritize action-layer defenses

   Increase investment in securing tool invocation, code execution sandboxes, and API call authorization, since the review shows these are underrepresented in current defense research relative to real-world impact.

2. 2

   Enforce layer isolation

   Architect agentic systems so that a compromise in the perception layer (e.g., prompt injection) cannot directly propagate to unchecked actions in the action layer; use strict boundaries, allow-lists, and human-in-the-loop gates for high-privilege operations.

3. 3

   Adopt the four-layer taxonomy for risk assessment

   Use the paper's perception/brain/action/interaction taxonomy to audit existing agent deployments and identify under-tested vulnerability classes such as tool misuse and sandbox escape.

4. 4

   Track emerging containment research

   Monitor follow-on work addressing the seven open problems on containment identified by the review, as these will likely inform future best practices for agent isolation.