hackagent 0.13.0__tar.gz → 0.13.1__tar.gz

{hackagent-0.13.0 → hackagent-0.13.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: hackagent
-Version: 0.13.0
+Version: 0.13.1
 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-0.13.0 → hackagent-0.13.1}/hackagent/attacks/evaluator/base.py

@@ -596,6 +596,11 @@ class BaseJudgeEvaluator(ABC):
             Tuple of (score, explanation).
         """
         # Step 1: Initial request
+        configured_thinking = getattr(self.config, "thinking", False)
+        if not isinstance(configured_thinking, bool):
+            configured_thinking = False
+        if request_data.get("thinking") is None:
+            request_data["thinking"] = configured_thinking
         response = self.agent_router.route_request(
             registration_key=self.agent_registration_key,
             request_data=request_data,
@@ -698,6 +703,11 @@ class BaseJudgeEvaluator(ABC):
         include_raw_response: bool = False,
     ) -> Tuple[Any, Optional[str]] | Tuple[Any, Optional[str], Optional[str]]:
         """Async equivalent of :meth:`_request_with_assertions`."""
+        configured_thinking = getattr(self.config, "thinking", False)
+        if not isinstance(configured_thinking, bool):
+            configured_thinking = False
+        if request_data.get("thinking") is None:
+            request_data["thinking"] = configured_thinking
         response = await self._route_request_async(
             registration_key=self.agent_registration_key,
             request_data=request_data,
@@ -967,4 +977,9 @@ Answer (yes or no):"""
             "messages": self._build_messages(feedback_prompt),
             "max_tokens": self.config.max_tokens_eval,
             "temperature": 0.0,  # Deterministic for retry
+            "thinking": (
+                original_request_data["thinking"]
+                if original_request_data.get("thinking") is not None
+                else False
+            ),
         }

{hackagent-0.13.0 → hackagent-0.13.1}/hackagent/attacks/evaluator/judge_evaluators.py

@@ -546,6 +546,11 @@ Do not output any other text.
             "messages": self._build_messages(feedback_prompt),
             "max_tokens": self.config.max_tokens_eval,
             "temperature": 0.0,
+            "thinking": (
+                original_request_data["thinking"]
+                if original_request_data.get("thinking") is not None
+                else False
+            ),
         }
 
 

{hackagent-0.13.0 → hackagent-0.13.1}/hackagent/attacks/techniques/advprefix/config.py

@@ -247,6 +247,9 @@ class EvaluatorConfig(BaseModel):
     max_judge_retries: int = 1
     # Optional user-defined system prompt prepended to every judge request.
     system_prompt: Optional[str] = None
+    # Reasoning control passed on every judge request. Missing/None is
+    # normalized to False by BaseJudgeEvaluator.
+    thinking: Optional[bool] = False
 
     @model_validator(mode="before")
     @classmethod

{hackagent-0.13.0 → hackagent-0.13.1}/hackagent/attacks/techniques/autodan_turbo/core.py

@@ -365,6 +365,7 @@ def score_response(
     max_retries=5,
     scorer_max_tokens=DEFAULT_MAX_OUTPUT_TOKENS,
     role_label="scorer",
+    thinking=False,
 ):
     """Score target output, using a wrapper only for unstructured assessments.
 
@@ -386,6 +387,8 @@ def score_response(
         Explicit assessment scores take precedence over model-based extraction.
         Exhausted retries return ``(1.0, "Score extraction failed")``.
     """
+    if not isinstance(thinking, bool):
+        thinking = False
     goal_text = str(goal or "")
     system = SCORER_SYSTEM_PROMPT.format(
         goal=goal_text,
@@ -410,6 +413,7 @@ def score_response(
                     "max_tokens": scorer_max_tokens,
                     "temperature": 0.7,
                     "top_p": 0.9,
+                    "thinking": thinking,
                 },
             )
             assessment = extract_response_content(resp, logger)
@@ -438,7 +442,12 @@ def score_response(
             ]
             resp2 = scorer_router.route_request(
                 registration_key=scorer_key,
-                request_data={"messages": msgs2, "max_tokens": 50, "temperature": 0.0},
+                request_data={
+                    "messages": msgs2,
+                    "max_tokens": 50,
+                    "temperature": 0.0,
+                    "thinking": thinking,
+                },
             )
             score_text = extract_response_content(resp2, logger)
             if score_text:
@@ -479,6 +488,7 @@ def score_response(
                     "messages": direct_numeric_msgs,
                     "max_tokens": 16,
                     "temperature": 0.0,
+                    "thinking": thinking,
                 },
             )
             direct_text = extract_response_content(resp3, logger)

{hackagent-0.13.0 → hackagent-0.13.1}/hackagent/attacks/techniques/autodan_turbo/lifelong.py

@@ -332,6 +332,11 @@ def execute(
                     max_retries=max_parse_retries,
                     scorer_max_tokens=scorer_max_tokens,
                     role_label=scorer_label,
+                    thinking=(
+                        _judge_cfg.get("thinking")
+                        if isinstance(_judge_cfg.get("thinking"), bool)
+                        else False
+                    ),
                 )
                 emit_phase_trace(
                     config,

{hackagent-0.13.0 → hackagent-0.13.1}/hackagent/attacks/techniques/autodan_turbo/warm_up.py

@@ -228,6 +228,11 @@ def execute(
                     max_retries=max_parse_retries,
                     scorer_max_tokens=scorer_max_tokens,
                     role_label=scorer_label,
+                    thinking=(
+                        _judge_cfg.get("thinking")
+                        if isinstance(_judge_cfg.get("thinking"), bool)
+                        else False
+                    ),
                 )
                 emit_phase_trace(
                     config,

{hackagent-0.13.0 → hackagent-0.13.1}/hackagent/attacks/techniques/indirect_prompt_injection/attack.py

@@ -1350,6 +1350,17 @@ class IndirectPromptInjectionAttack(BaseAttack):
         self, query: str, context: str, response: str, goal: str
     ) -> Dict[str, Any]:
         """Use judge LLM to evaluate if target followed poisoned instructions."""
+        judge_config = (
+            self.config.get("judges", [{}])[0]
+            if self.config.get("judges")
+            else self.config.get("judge", {})
+        )
+        configured_thinking = (
+            judge_config.get("thinking") if isinstance(judge_config, dict) else None
+        )
+        thinking = (
+            configured_thinking if isinstance(configured_thinking, bool) else False
+        )
         user_message = (
             f"BENIGN USER QUERY:\n{query}\n\n"
             f"RETRIEVED CONTEXT:\n{context[:3000]}\n\n"
@@ -1366,6 +1377,7 @@ class IndirectPromptInjectionAttack(BaseAttack):
                 ],
                 "max_tokens": 500,
                 "temperature": 0.0,
+                "thinking": thinking,
             },
         )
 

{hackagent-0.13.0 → hackagent-0.13.1}/hackagent/attacks/techniques/pair/attack.py

@@ -726,6 +726,11 @@ class PAIRAttack(BaseAttack):
                 max_retries=max_parse_retries,
                 scorer_max_tokens=scorer_max_tokens,
                 role_label="pair-judge",
+                thinking=(
+                    judge_config.get("thinking")
+                    if isinstance(judge_config.get("thinking"), bool)
+                    else False
+                ),
             )
             self._set_scorer_explanation(assessment or "")
 
@@ -812,6 +817,13 @@ class PAIRAttack(BaseAttack):
 
         judge_prompt_max = int(self.config.get("judge_prompt_max_chars", 2500))
         judge_response_max = int(self.config.get("judge_response_max_chars", 3500))
+        judge_config = self.config.get("judge") or self.config.get("scorer", {})
+        configured_thinking = (
+            judge_config.get("thinking") if isinstance(judge_config, dict) else None
+        )
+        thinking = (
+            configured_thinking if isinstance(configured_thinking, bool) else False
+        )
 
         prompt_for_judge = _truncate_text(prompt or "", judge_prompt_max)
         response_for_judge = _truncate_text(response or "", judge_response_max)
@@ -833,7 +845,6 @@ class PAIRAttack(BaseAttack):
         ]
 
         try:
-            judge_config = self.config.get("judge") or self.config.get("scorer", {})
             request_timeout = judge_config.get(
                 "timeout",
                 judge_config.get("request_timeout", self.config.get("timeout", 120)),
@@ -848,6 +859,7 @@ class PAIRAttack(BaseAttack):
                     "max_tokens": max_tokens,
                     "temperature": 0.0,
                     "timeout": request_timeout,
+                    "thinking": thinking,
                 },
             )
             legacy_text = extract_response_content(legacy_response, self.logger)

{hackagent-0.13.0 → hackagent-0.13.1}/hackagent/attacks/techniques/rag/attack.py

@@ -1382,6 +1382,17 @@ class RagAttack(BaseAttack):
         self, query: str, context: str, response: str, goal: str
     ) -> Dict[str, Any]:
         """Use judge LLM to evaluate if target followed poisoned instructions."""
+        judge_config = (
+            self.config.get("judges", [{}])[0]
+            if self.config.get("judges")
+            else self.config.get("judge", {})
+        )
+        configured_thinking = (
+            judge_config.get("thinking") if isinstance(judge_config, dict) else None
+        )
+        thinking = (
+            configured_thinking if isinstance(configured_thinking, bool) else False
+        )
         user_message = (
             f"BENIGN USER QUERY:\n{query}\n\n"
             f"RETRIEVED CONTEXT:\n{context[:3000]}\n\n"
@@ -1398,6 +1409,7 @@ class RagAttack(BaseAttack):
                 ],
                 "max_tokens": 500,
                 "temperature": 0.0,
+                "thinking": thinking,
             },
         )
 

{hackagent-0.13.0 → hackagent-0.13.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hackagent"
-version = "0.13.0"
+version = "0.13.1"
 description = "HackAgent is an open-source security toolkit to detect vulnerabilities of your AI Agents."
 authors = [
     {name = "AI Security Lab", email = "ais@ai4i.it"}