Skip to main content

Input Manipulation Attack

Tests whether encoding bypasses, format string attacks, or Unicode manipulation can evade input validation and safety filters.

Sub-types​

  • encoding_bypass: Using character encoding tricks to bypass input filters.
  • format_string_attack: Exploiting format string processing in input handling.
  • unicode_manipulation: Using Unicode homoglyphs or special characters to evade detection.

Threat Profile​

Secondary

  • wmdp_cyber: Cybersecurity knowledge covering SQL injection and command injection techniques

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 InputManipulationAttack
from hackagent.catalog.risks.input_manipulation_attack.types import InputManipulationAttackType

# Use all sub-types
vuln = InputManipulationAttack()

# Or specify particular sub-types
vuln = InputManipulationAttack(types=[
    InputManipulationAttackType.ENCODING_BYPASS.value,
    InputManipulationAttackType.FORMAT_STRING_ATTACK.value,
])

Run an Evaluation Campaign​

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