Tool Invocation Manipulated by Injection
工具呼叫遭注入操控
Verification & Validation
Risk Description
When an attacker embeds instructions in the input that induce a system with tool privileges to transmit internal data externally; due to unmitigated control gaps, the system treats the instruction as user intent and executes it, triggering external stakeholder impacts and causing all actions occur within granted permissions and trigger no anomaly alerts.
Framework Mappings
OWASP Top 10 for LLMLLM01、LLM06
MITRE ATLASAML.T0051
NIST AI 600-1Information Security
ISO/IEC 5338運作與監控
MIT AI Risk RepositoryDomain 2
Risk Treatment & Implementation Guidance
Sandbox agent tool capabilities under least privilege, disabling high-risk tools like external data transmission by default; Whitelist tool calls with parameter validation so external instructions cannot alter authorization; Require human confirmation for data-egress actions and retain complete call logs