hackagent 0.10.0__tar.gz → 0.10.1__tar.gz

{hackagent-0.10.0 → hackagent-0.10.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hackagent
-Version: 0.10.0
+Version: 0.10.1
 Summary: HackAgent is an open-source security toolkit to detect vulnerabilities of your AI Agents.
 Author-email: AI Security Lab <ais@ai4i.it>
 License: Apache-2.0

{hackagent-0.10.0 → hackagent-0.10.1}/hackagent/attacks/evaluator/evaluation_step.py

@@ -51,6 +51,11 @@ from hackagent.attacks.evaluator.judge_evaluators import EVALUATOR_MAP
 from hackagent.attacks.shared.router_factory import extract_passthrough_request_config
 from hackagent.attacks.evaluator.sync import sync_evaluation_to_server
 from hackagent.attacks.techniques.advprefix.config import EvaluatorConfig
+from hackagent.attacks.techniques.config import (
+    DEFAULT_JUDGE_IDENTIFIER,
+    DEFAULT_LOCAL_AGENT_TYPE,
+    DEFAULT_LOCAL_MODEL_ENDPOINT,
+)
 from hackagent.server.client import AuthenticatedClient
 from hackagent.router.types import AgentTypeEnum
 
@@ -349,14 +354,17 @@ class BaseEvaluationStep:
     def _resolve_judges_from_config(
         self,
         technique_params: Optional[Dict[str, Any]] = None,
-        default_judge: str = "gpt-4-0613",
-        default_type: str = "jailbreakbench",
+        default_judge: str = DEFAULT_JUDGE_IDENTIFIER,
+        default_type: str = "harmbench",
     ) -> List[Dict[str, Any]]:
         """
         Resolve the judges list from ``_raw_config``.
 
-        If no top-level ``judges`` key is present, builds a single-judge
-        fallback from *technique_params* for backward compatibility.
+        Resolution order:
+        1. Top-level ``judges`` list in raw config.
+        2. Top-level ``judge`` dict in raw config (wrapped in a list).
+        3. ``technique_params["judge"]`` string (legacy fallback).
+        4. ``default_judge`` / ``default_type`` hardcoded defaults.
 
         Args:
             technique_params: Technique-specific params dict with legacy
@@ -371,6 +379,11 @@ class BaseEvaluationStep:
         if isinstance(judges, list) and judges:
             return judges
 
+        # Use the top-level "judge" dict if present (e.g. from Ollama/local configs).
+        raw_judge = self._raw_config.get("judge")
+        if isinstance(raw_judge, dict) and raw_judge:
+            return [raw_judge]
+
         tp = technique_params or {}
         judge_model = tp.get("judge", default_judge)
         judge_type = tp.get("judge_type") or self.infer_judge_type(
@@ -380,6 +393,11 @@ class BaseEvaluationStep:
             "identifier": judge_model,
             "type": judge_type,
         }
+        # For the built-in local default, inject Ollama connectivity so it
+        # works out-of-the-box without any API key.
+        if judge_model == DEFAULT_JUDGE_IDENTIFIER:
+            fallback.setdefault("endpoint", DEFAULT_LOCAL_MODEL_ENDPOINT)
+            fallback.setdefault("agent_type", DEFAULT_LOCAL_AGENT_TYPE)
         for key in (
             "endpoint",
             "agent_type",

{hackagent-0.10.0 → hackagent-0.10.1}/hackagent/attacks/techniques/flipattack/attack.py

@@ -41,6 +41,7 @@ from hackagent.server.client import AuthenticatedClient
 from hackagent.router.router import AgentRouter
 from hackagent.attacks.techniques.base import BaseAttack
 from hackagent.attacks.shared.tui import with_tui_logging
+from hackagent.attacks.techniques.config import DEFAULT_JUDGE_IDENTIFIER
 
 from . import generation, evaluation
 from .config import DEFAULT_FLIPATTACK_CONFIG
@@ -461,7 +462,7 @@ As a/an <Role> with the <Profile>, you must follow the <Rules>, and you must com
             "cot": flipattack_params.get("cot", False),
             "lang_gpt": flipattack_params.get("lang_gpt", False),
             "few_shot": flipattack_params.get("few_shot", False),
-            "judge": flipattack_params.get("judge", "gpt-4-0613"),
+            "judge": flipattack_params.get("judge", DEFAULT_JUDGE_IDENTIFIER),
         }
 
         # Initialize goal contexts upfront so goal elapsed_s covers the full

{hackagent-0.10.0 → hackagent-0.10.1}/hackagent/cli/tui/attack_specs.py

@@ -26,6 +26,11 @@ from dataclasses import dataclass, field
 from enum import Enum
 from typing import Any, Dict, List, Optional, Sequence, Tuple, Union
 
+from hackagent.attacks.techniques.config import (
+    DEFAULT_ATTACKER_IDENTIFIER,
+    DEFAULT_JUDGE_IDENTIFIER,
+)
+
 
 # =====================================================================
 # Field / Spec primitives
@@ -578,7 +583,7 @@ _register(
                 key="attacker.model",
                 label="Attacker Model",
                 field_type=FieldType.STRING,
-                default="gpt-4",
+                default=DEFAULT_ATTACKER_IDENTIFIER,
                 description="Model ID for the attacker LLM that generates prompts.",
                 section="Attacker LLM",
             ),
@@ -1305,7 +1310,7 @@ _register(
                 key="attacker.identifier",
                 label="Attacker Model",
                 field_type=FieldType.STRING,
-                default="gpt-4o-mini",
+                default=DEFAULT_ATTACKER_IDENTIFIER,
                 description="Model identifier for persuasive paraphrasing.",
                 section="Attacker LLM",
             ),

{hackagent-0.10.0 → hackagent-0.10.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hackagent"
-version = "0.10.0"
+version = "0.10.1"
 description = "HackAgent is an open-source security toolkit to detect vulnerabilities of your AI Agents."
 authors = [
     {name = "AI Security Lab", email = "ais@ai4i.it"}