pres_security_benchmarking_llm/pages/advanced-att-techniques.md

Advanced Attack Techniques

Prompt Obfuscation

Using techniques like Base64 encoding, character transformations (e.g., ROT13), or prompt-level obfuscations to bypass restrictions.

Model-based Jailbreaking

Automating the creation of adversarial attacks by evolving simple synthetic inputs into more complex attacks.

Dialogue-based Jailbreaking

Employing reinforcement learning with two models: the target LLM and a red-teamer model trained to exploit vulnerabilities.

Primary Areas of Concern

• Organizational reputation damage
• Legal compliance violations
• Data security breaches