Attack Algorithms
DarkCite
Authority citation-driven jailbreak technique that leverages models' inherent bias toward authority by including authoritative citations, matching optimal citation types to specific risk categories, and generating credible but fabricated citations relevant to harmful instructions.
Bijection Learning
Language bijection jailbreak technique that teaches models a custom 'Language Alpha' using bijection mappings (digit or character), provides multiple teaching examples to help the model learn the mapping, and asks harmful prompts in the transformed language to bypass safety filters.
Crescendo Attack
Multi-turn jailbreak strategy that uses an attack model to generate a series of questions to progressively guide the target model to generate harmful content, with backtracking when refusals are encountered.
Flip Attack
Sophisticated jailbreak technique that disguises harmful prompts using text flipping (words, characters, sentences), uses guidance modules (CoT, LangGPT, few-shot) to teach models to denoise, and achieves high success rates with single queries.
Language Game
Language encoding jailbreak attack that encodes harmful prompts using various language transformations (leet speak, pig latin, etc.), instructs the model to respond in the encoded format, and uses custom formatting rules to disguise harmful intent.
BoN Attack
Text augmentation-based jailbreak strategy that applies multiple transformations (scrambling, capitalization, ASCII perturbations) to harmful prompts, tests each variant against the target model, and selects the most effective candidate based on Attack Success Rate.
Character-Level Attacks
Fine-grained manipulations at the token or symbol level to evade detection.
Optimization-Based Attacks
Algorithmically searching for the most effective jailbreak or harmful prompt.
Multi-Turn & Contextual Attacks
Exploiting conversation history or memory to bypass protections.
Poisoning Attacks
Contaminating training or fine-tuning data to degrade safety.
Sentence-Level Attacks
Semantic rewrites and contextual manipulations to induce unsafe responses.
Decoding-Time Manipulation
Influencing model output through decoding strategies or generation parameters.
Role & Identity Manipulation
Convincing the model to take on privileged or unsafe roles.
Policy-Evasion Attacks
Targeting guardrail logic gaps or policy inconsistencies.
Cross-Modal Attacks (multimodal models)
Using one modality to inject harmful input into another.
Tool-Use Exploitation (agentic systems)
Manipulating APIs, plugins, or external tools for harmful actions.