S0118RS3-T10-S0118-Z · Full risk code

Direct Prompt Injection Bypassing Safeguards

直接提示注入繞過防護

Verification & Validation
Risk Description

When a malicious actor conducts direct prompt injection attacks via multi-turn dialogue and role-playing; because system safeguards possess only single-turn semantic checking without cross-turn intent analysis, the attacker successfully bypasses safety alignment guardrails, leading to unlawful output and compliance liability.

Framework Mappings

OWASP Top 10 for LLMLLM01
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

Deploy prompt-injection detection with cross-turn intent analysis for progressive elicitation rather than single-turn review; Guard the output layer so final outputs still pass content-policy checks; Set cumulative risk thresholds on sensitive topics, escalating when gradual multi-turn approaches trip them