Skip to main content
LLM Security Database
Skip to research details
Back to research findings
LMVD-ID: 79625393
Paper published July 21, 2026
Entry analyzed September 9, 2026
Paper-reported evidence
Confidence: Source-linked

The LMVD-ID is an internal research identifier, not an official CVE identifier.

Unsafe shell commands can pass LLM agent dispatch guards

Shell-command screening can miss harmful actions generated by coding agents. CARE combines static checks with optional model review at the command-dispatch boundary.

Application layer
Prompt layer
Injection
AI agents
Coding agents
Data security
Integrity
Safety
Read primary paperBibTeX citation

Paper-evaluated models(4)

  • Qwen3-Coder-30B-A3B-Instruct
  • Granite Guardian 2B
  • Llama Guard 3 8B
  • GPT-oss Safeguard 20B

Description

Shell-command screening can miss harmful actions generated by coding agents. CARE combines static checks with optional model review at the command-dispatch boundary.

Examples

See the primary evaluation (opens in a new tab).

Impact

Across 600 author-generated commands in Docker, harmful execution falls from 74.83% without screening to 37.33% with static CARE; adding model review raises it to 52.50%. This measures individual commands, not host containment.

Affected Systems

  • Shell-executing coding agents and CARE command guards.

Mitigation Steps

  • Check commands before dispatch and retain operating-system isolation.
  • Measure false positives and harmful execution separately.
  • Validate model overrides of static decisions on representative workloads.

Evidence

Research context and confidence

Evidence and verification
Paper-reported; independent reproduction is not documented.
Primary source plus a dedicated evidence section.
Severity
Not rated by this catalog.
Source and publication type
arXiv · Research preprint.
Peer-review status is not provided by this source.
Author and publication status
Author metadata is not stored; see the primary paper.
Threat model and attacker access
Ability to influence untrusted model inputs or connected content.
Related deployment categories
Agent workflows; Coding agents
Taxonomy labels only; paper-specific deployment prerequisites are not inferred.
Affected systems
Shell-executing coding agents and CARE command guards.

Research Paper

CARE: Pre-Execution Command Verification for Shell-Executing LLM Agents

Primary source: arXiv. Findings are reported by the cited research and have not been independently verified.

View Paper

Related research