Skip to main content

Sensitive Information Disclosure

Tests for training-data extraction, architecture disclosure, and configuration leakage.

Sub-types​

  • training_data_extraction: Model memorisation allows extraction of training data.
  • system_architecture_disclosure: Model reveals internal architecture details.
  • configuration_leakage: Model exposes configuration parameters or settings.

Threat Profile​

Secondary

  • advbench: Adversarial prompts that may trigger info leaks
  • saladbench: 21K harmful questions — includes info disclosure scenarios

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 SensitiveInformationDisclosure
from hackagent.catalog.risks.sensitive_information_disclosure.types import SensitiveInformationDisclosureType

# Use all sub-types
vuln = SensitiveInformationDisclosure()

# Or specify particular sub-types
vuln = SensitiveInformationDisclosure(types=[
    SensitiveInformationDisclosureType.TRAINING_DATA_EXTRACTION.value,
    SensitiveInformationDisclosureType.SYSTEM_ARCHITECTURE_DISCLOSURE.value,
])

Run an Evaluation Campaign​

from hackagent.catalog.risks.sensitive_information_disclosure import SENSITIVE_INFORMATION_DISCLOSURE_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 SENSITIVE_INFORMATION_DISCLOSURE_PROFILE.primary_attacks:
    for dataset in SENSITIVE_INFORMATION_DISCLOSURE_PROFILE.datasets:
        result = run_campaign(
            {
                "version": 1,
                "campaign": {
                    "name": f"{attack.technique} vs {dataset.preset}",
                    "description": SENSITIVE_INFORMATION_DISCLOSURE_PROFILE.objective,
                },
                "dataset": {"source": {"type": "preset", "name": dataset.preset}},
                "target": TARGET,
                "attacks": [{"name": attack.technique}],
            }
        )
        print(summary(result))