Privilege escalation with prompt injection
Privilege escalation attacks in LLMs that use prompt injection largely depend on the LLM’s integration with other systems. The risk of such attacks is relatively low in standalone LLMs but significantly higher when LLMs are integrated with external systems with APIs or endpoints:
- Downstream systems vulnerabilities: In scenarios where LLMs are integrated with downstream systems, prompt injection can be used to manipulate these systems. For example, using LangChain’s
create_sql_query_chain, attackers can generate Data Manipulation Language (DML) statements, leading to unauthorized data deletion or manipulation. This also includes the unauthorized sending of emails or messages. OWASP LLM08 (Excessive Agency: https://genai.owasp.org/llmrisk/llm08-excessive-agency/) addresses this, noting the risks of unintentional or accidental privilege escalation if LLMs are granted excessive functionality or access rights. - Plugin-based...
