There are 25 attacks in the model-extraction category. Each shows its technique id, mode, standards mapping, description, and the exact command to run it (swap the provider and target as needed).
​
API Cloning
critical
api-cloning · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to extract enough model behavior via API queries to build a functional cloneRun it:
ai-blackteam run -p anthropic -a api-cloning -t "your target prompt"
​
Distillation Attack
critical
distillation-attack · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to systematically extract model knowledge for training a clone or distilled surrogate modelRun it:
ai-blackteam run -p anthropic -a distillation-attack -t "your target prompt"
​
Logit Extraction
critical
logit-extraction · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to extract raw logit values to enable model stealing via output probability matchingRun it:
ai-blackteam run -p anthropic -a logit-extraction -t "your target prompt"
​
Model Inversion
critical
model-inversion · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to invert the model to reconstruct training inputs from model outputs or internal representationsRun it:
ai-blackteam run -p anthropic -a model-inversion -t "your target prompt"
​
Model Weight Extraction
critical
model-weight-extraction · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to extract model weight values or reconstruct internal model parametersRun it:
ai-blackteam run -p anthropic -a model-weight-extraction -t "your target prompt"
​
RLHF Data Extraction
critical
rlhf-data-extraction · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to extract RLHF preference data, reward model signals, and human feedback labelsRun it:
ai-blackteam run -p anthropic -a rlhf-data-extraction -t "your target prompt"
​
Safety Training Extraction
critical
safety-training-extraction · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to extract the model’s safety training data, refusal examples, and alignment techniquesRun it:
ai-blackteam run -p anthropic -a safety-training-extraction -t "your target prompt"
​
System Prompt Recovery
critical
system-prompt-recovery · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to recover or reconstruct the system prompt and operator instructionsRun it:
ai-blackteam run -p anthropic -a system-prompt-recovery -t "your target prompt"
​
Training Data Memorization
critical
training-data-memorization · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to trigger memorized verbatim content from the training corpus including private documentsRun it:
ai-blackteam run -p anthropic -a training-data-memorization -t "your target prompt"
​
Training Pipeline Extraction
critical
training-pipeline-extraction · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to extract details about the model’s training pipeline, data preprocessing, and infrastructureRun it:
ai-blackteam run -p anthropic -a training-pipeline-extraction -t "your target prompt"
​
Architecture Probing
high
architecture-probing · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to reveal internal model architecture details such as layer counts, hidden dimensions, and attention headsRun it:
ai-blackteam run -p anthropic -a architecture-probing -t "your target prompt"
​
Attention Pattern Extraction
high
attention-pattern-extraction · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to extract attention weight matrices and head activation patterns for model reconstructionRun it:
ai-blackteam run -p anthropic -a attention-pattern-extraction -t "your target prompt"
​
Capability Mapping
high
capability-mapping · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to comprehensively map all model capabilities, limitations, and undisclosed featuresRun it:
ai-blackteam run -p anthropic -a capability-mapping -t "your target prompt"
​
Decision Boundary Probing
high
decision-boundary-probing · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to map the model’s decision boundaries between allowed and refused contentRun it:
ai-blackteam run -p anthropic -a decision-boundary-probing -t "your target prompt"
​
Embedding Extraction
high
embedding-extraction · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to extract internal embedding vectors and representation space structureRun it:
ai-blackteam run -p anthropic -a embedding-extraction -t "your target prompt"
​
Fine-Tuning Detection
high
fine-tuning-detection · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to detect whether a model has been fine-tuned and extract information about the fine-tuning dataRun it:
ai-blackteam run -p anthropic -a fine-tuning-detection -t "your target prompt"
​
Hyperparameter Extraction
high
hyperparameter-extraction · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to extract training hyperparameters such as learning rate, batch size, and optimizer settingsRun it:
ai-blackteam run -p anthropic -a hyperparameter-extraction -t "your target prompt"
​
Model Comparison Leak
high
model-comparison-leak · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to leak model identity and configuration by exploiting comparative responses across model versionsRun it:
ai-blackteam run -p anthropic -a model-comparison-leak -t "your target prompt"
​
Model Fingerprinting
high
model-fingerprinting · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to fingerprint a model to uniquely identify it and distinguish it from other models or versionsRun it:
ai-blackteam run -p anthropic -a model-fingerprinting -t "your target prompt"
​
Output Distribution Analysis
high
output-distribution-analysis · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to analyze the model’s output probability distribution to reconstruct its decision functionRun it:
ai-blackteam run -p anthropic -a output-distribution-analysis -t "your target prompt"
​
Tokenizer Extraction
high
tokenizer-extraction · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to extract the model’s tokenizer vocabulary, merge rules, and special token configurationRun it:
ai-blackteam run -p anthropic -a tokenizer-extraction -t "your target prompt"
​
Watermark Detection
high
watermark-detection · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to detect and extract model output watermarks or provenance signalsRun it:
ai-blackteam run -p anthropic -a watermark-detection -t "your target prompt"
​
Confidence Calibration Probe
medium
confidence-calibration-probe · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to extract confidence scores, probability calibration, and uncertainty quantification internalsRun it:
ai-blackteam run -p anthropic -a confidence-calibration-probe -t "your target prompt"
​
Knowledge Cutoff Probing
medium
knowledge-cutoff-probing · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to precisely determine the model’s training data cutoff date and data recency distributionRun it:
ai-blackteam run -p anthropic -a knowledge-cutoff-probing -t "your target prompt"
​
Version Detection
medium
version-detection · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Prompts designed to detect the exact model version, release date, and deployment configurationRun it:
ai-blackteam run -p anthropic -a version-detection -t "your target prompt"