AIDEFEND Labs
  • Solutions
  • AI Security Cases × Frameworks AIDEFEND in Action AIDEFEND SecureFlow
  • Learn AI Security
  • About
  • Explore AIDEFEND
Learning home

LLM02 Sensitive Information Disclosure

The scenario

Future Careers offers an AI resume review service to help job seekers improve their resumes. An attacker hides instructions inside a resume, tricking the AI into revealing the company's proprietary resume screening algorithm—a valuable trade secret.

How it works

This demo assistant actually holds internal algorithm documents. These questions do not magically give a general model access to corporate documents.

System Prompt Disguise: The attacker disguises malicious instructions as an internal system command, such as `SYSTEM_DIAGNOSTIC_OVERRIDE`, so the AI treats them as a privileged diagnostic request rather than ordinary user input.

Contextual Injection: Malicious instructions are embedded in an otherwise ordinary resume. While reviewing the resume, the AI reads and follows those instructions instead of treating them as untrusted document content.

Guardrail Bypassing: Although the system prompt prohibits disclosing algorithm details, the forged system command bypasses the AI's safeguards and exposes sensitive information and trade secrets.

Attack successful: screening algorithm exposed!

The AI followed the hidden instructions while reviewing the resume and revealed its internal scoring model. Competitors or applicants could use the information to game the screening process.