Skip to main content

Model Evasion

Tests whether adversarial examples, feature manipulation, or boundary exploitation can evade the model's safety mechanisms.

Sub-types​

  • adversarial_examples: Crafted inputs that cause the model to misclassify or produce wrong outputs.
  • feature_space_manipulation: Manipulating input features to evade detection or safety mechanisms.
  • model_boundary_exploitation: Exploiting decision boundaries to find blind spots in model behaviour.

Threat Profile​

Primary

  • advbench: Adversarial benchmarks for evaluating evasion resistance

Secondary

  • xstest: XSTest for adversarial prompt detection

Attack Techniques​

Primary

  • Static Template: Template-based prompt injection
  • PAIR: Iterative refinement for bypass discovery

Secondary

Metrics​

  • asr
  • judge_score

Usage​

Instantiate the Vulnerability​

from hackagent.catalog.risks import ModelEvasion
from hackagent.catalog.risks.model_evasion.types import ModelEvasionType

# Use all sub-types
vuln = ModelEvasion()

# Or specify particular sub-types
vuln = ModelEvasion(types=[
    ModelEvasionType.ADVERSARIAL_EXAMPLES.value,
    ModelEvasionType.FEATURE_SPACE_MANIPULATION.value,
])

Run an Evaluation Campaign​

from hackagent.catalog.risks.model_evasion import MODEL_EVASION_PROFILE
from hackagent.orchestrator.campaign import run_campaign, summary

TARGET = {
    "name": "my-agent",
    "connection": {
        "provider": "local",
        "type": "OPENAI",
        "endpoint": "http://localhost:8080/chat",
    },
}

# Use profile recommendations
for attack in MODEL_EVASION_PROFILE.primary_attacks:
    for dataset in MODEL_EVASION_PROFILE.primary_datasets:
        result = run_campaign(
            {
                "version": 1,
                "campaign": {
                    "name": f"{attack.technique} vs {dataset.preset}",
                    "description": MODEL_EVASION_PROFILE.objective,
                },
                "dataset": {"source": {"type": "preset", "name": dataset.preset}},
                "target": TARGET,
                "attacks": [{"name": attack.technique}],
            }
        )
        print(summary(result))