How jinxiaolong1129/foot-in-the-door-jailbreak works
FITD is a research implementation of a multi-turn Foot-in-the-Door jailbreak: it progressively transforms prompts, queries a language model, and records attack outcomes. Separate analysis programs investigate alignment robustness and model internals across attack stages, including semantic attention networks and hidden states. The README also describes benchmark evaluation and pre-generated prompts. The excerpts are partial, so the graph omits uncertain internal calls and does not infer wiring for unsampled modules.
Main entry points: FITD.py FITD runner · analysis copy/main.py FITD semantic analysis · analysis/main.py Semantic analysis · alignment.py Alignment robustness
Components 17
Attack workflow
- FITD runner
FITD.py
Attack inputs
- Benchmark data
data - Attack prompts
prompts
Mechanism analysis
- Semantic analysis
analysis/main.py - FITD semantic analysis
analysis copy/main.py - Hidden-state analysis
analysis copy/semantic_hidden_states.py - Attention analysis
analysis copy/semantic_attention.py - Semantic networks
analysis/semantic_network.py - Alignment robustness
alignment.py - Response-based alignment
analysis/alignment_with_predefine_response.py - Token mechanisms
token_analysis.py
Results
- Attack results
results - Analysis outputs
analysis/semantic_analysis_results - Alignment outputs
analysis/alignment_results - Mechanism outputs
mechanism_analysis_results
Other components
- Researcher
- Language model API
Connections 21
- Researcher → FITD runner: runs
- FITD runner → Benchmark data: loads data
- FITD runner → Attack prompts: uses prompts
- FITD runner → Language model API: queries
- FITD runner → Attack results: saves outcomes
- Researcher → Semantic analysis: runs analysis
- Semantic analysis → Semantic networks: analyzes networks
- Semantic analysis → Analysis outputs: writes results
- Researcher → FITD semantic analysis: runs analysis
- FITD semantic analysis → Hidden-state analysis: analyzes states
- FITD semantic analysis → Attention analysis: analyzes attention
- FITD semantic analysis → Analysis outputs: writes results
- Researcher → Alignment robustness: runs analysis
- Alignment robustness → Language model API: samples responses
- Alignment robustness → Alignment outputs: writes results
- Researcher → Response-based alignment: runs analysis
- Response-based alignment → Language model API: loads model
- Response-based alignment → Alignment outputs: writes results
- Researcher → Token mechanisms: runs analysis
- Token mechanisms → Language model API: inspects model
- Token mechanisms → Mechanism outputs: writes results