Skip to main content

Excessive Agency

Tests whether the LLM performs actions or grants permissions exceeding its intended scope without oversight.

Sub-types​

  • functionality: Model performs actions beyond its intended scope.
  • permissions: Model grants or exercises permissions it should not have.
  • autonomy: Model acts autonomously without required human oversight.

Threat Profile​

Primary

  • agentharm: AgentHarm dataset for evaluating excessive agency in tool use

Secondary

  • agentharm_benign: Benign agent scenarios to establish baseline behavior

Attack Techniques​

Primary

Metrics​

  • asr
  • judge_score

Usage​

Instantiate the Vulnerability​

from hackagent.catalog.risks import ExcessiveAgency
from hackagent.catalog.risks.excessive_agency.types import ExcessiveAgencyType

# Use all sub-types
vuln = ExcessiveAgency()

# Or specify particular sub-types
vuln = ExcessiveAgency(types=[
    ExcessiveAgencyType.FUNCTIONALITY.value,
    ExcessiveAgencyType.PERMISSIONS.value,
])

Run an Evaluation Campaign​

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