Merge pull request #47 from GitHubSecurityLab/filter_severity

add taskflow to filter severity

Author: m-y-mo

src/seclab_taskflows/mcp_servers/repo_context.py

@@ -15,6 +15,7 @@
 from seclab_taskflow_agent.path_utils import mcp_data_dir, log_file_name
 
 from .repo_context_models import Application, EntryPoint, UserAction, WebEntryPoint, ApplicationIssue, AuditResult, Base
+from .repo_context_models import LowSeverityAuditResult
 from .utils import process_repo
 
 logging.basicConfig(
@@ -26,7 +27,6 @@
 
 MEMORY = mcp_data_dir("seclab-taskflows", "repo_context", "REPO_CONTEXT_DIR")
 
-
 def app_to_dict(result):
     return {
         "app_id": result.id,
@@ -107,6 +107,7 @@ def __init__(self, memcache_state_dir: str):
                 WebEntryPoint.__table__,
                 ApplicationIssue.__table__,
                 AuditResult.__table__,
+                LowSeverityAuditResult.__table__,
             ],
         )
 
@@ -239,6 +240,22 @@ def store_new_user_action(self, repo, app_id, file, line, notes, update=False):
                 session.add(new_user_action)
             session.commit()
         return f"Updated or added user action for {file} and {line} in {repo}."
+    
+    def store_low_severity_reason(self, repo, component_id, result_id, reason):
+        with Session(self.engine) as session:
+            existing = session.query(LowSeverityAuditResult).filter_by(repo=repo, result_id=result_id).first()
+            if existing:
+                existing.reason += reason
+            else:
+                new_low_severity_result = LowSeverityAuditResult(
+                    repo=repo,
+                    component_id=component_id,
+                    result_id=result_id,
+                    reason=reason,
+                )
+                session.add(new_low_severity_result)
+            session.commit()
+        return f"Updated or added low severity result for {repo} and result id {result_id}"
 
     def get_app(self, repo, location):
         with Session(self.engine) as session:
@@ -294,6 +311,7 @@ def get_app_audit_results(self, repo, component_id, has_non_security_error, has_
                 "repo": app.repo,
                 "issue_type": issue.issue_type,
                 "issue_id": issue.issue_id,
+                "result_id" : issue.id,
                 "notes": issue.notes,
                 "has_vulnerability": issue.has_vulnerability,
                 "has_non_security_error": issue.has_non_security_error,
@@ -389,6 +407,7 @@ def clear_repo(self, repo):
             session.query(ApplicationIssue).filter_by(repo=repo).delete()
             session.query(WebEntryPoint).filter_by(repo=repo).delete()
             session.query(AuditResult).filter_by(repo=repo).delete()
+            session.query(LowSeverityAuditResult).filter_by(repo=repo).delete()
             session.commit()
         return f"Cleared results for repo {repo}"
 
@@ -782,6 +801,19 @@ def get_potential_audit_results_for_repo(
         backend.get_app_audit_results(repo, component_id=None, has_non_security_error=True, has_vulnerability=None)
     )
 
+@mcp.tool()
+def store_low_severity_reason(
+    owner: str = Field(description="The owner of the GitHub repository"),
+    repo: str = Field(description="The name of the GitHub repository"),
+    component_id: int = Field(description="The ID of the component"),
+    result_id: int = Field(description="The ID of the audit result"),
+    reason: str = Field(description="The reason why this issue is not considered high severity"),
+):
+    """
+    Store the reason for auditing an issue as low severity.
+    """
+    repo = process_repo(owner, repo)
+    return backend.store_low_severity_reason(repo, component_id, result_id, reason)
 
 @mcp.tool()
 def clear_repo(

src/seclab_taskflows/mcp_servers/repo_context_models.py

@@ -60,6 +60,19 @@ def __repr__(self):
             f"component_id={self.component_id}, issue_type={self.issue_type}, issue_id={self.issue_id}, notes={self.notes})>"
         )
 
+class LowSeverityAuditResult(Base):
+    __tablename__ = "low_severity_audit_result"
+    id: Mapped[int] = mapped_column(primary_key=True)
+    repo: Mapped[str]
+    component_id = Column(Integer, ForeignKey("application.id", ondelete="CASCADE"))
+    result_id = Column(Integer, ForeignKey("audit_result.id", ondelete="CASCADE"))
+    reason: Mapped[str] = mapped_column(Text)
+
+    def __repr__(self):
+        return (
+            f"<LowSeverityAuditResult(id={self.id}, repo={self.repo}, "
+            f"component_id={self.component_id}, result_id={self.result_id}, reason={self.reason})>"
+        )
 
 class EntryPoint(Base):
     __tablename__ = "entry_point"

src/seclab_taskflows/taskflows/audit/filter_severity.yaml

@@ -0,0 +1,73 @@
+# SPDX-FileCopyrightText: 2025 GitHub
+# SPDX-License-Identifier: MIT
+
+seclab-taskflow-agent:
+  filetype: taskflow
+  version: "1.0"
+
+model_config: seclab_taskflows.configs.model_config
+
+globals:
+  repo:
+taskflow:
+  - task:
+      must_complete: true
+      exclude_from_context: true
+      agents:
+        - seclab_taskflow_agent.personalities.assistant
+      model: general_tasks
+      user_prompt: |
+        Fetch the vulnerable issues in {{ globals.repo }}.
+      toolboxes:
+        - seclab_taskflows.toolboxes.repo_context
+  - task:
+      must_complete: true
+      repeat_prompt: true
+      async: true
+      agents:
+        - seclab_taskflows.personalities.web_application_security_expert
+      model: code_analysis
+      toolboxes:
+        - seclab_taskflows.toolboxes.repo_context
+        - seclab_taskflows.toolboxes.local_file_viewer
+      user_prompt: |
+        The audit result has id {{ result.result_id }}. It is in the component with id {{ result.component_id }}
+        at the location {{ result.location }}. It is in the repository {{ result.repo }}. 
+        The notes for the audit result are as follows:
+        
+        {{ result.notes }}
+
+        Determine the low severity results based on the following criteria:
+        - blind SSRF that has limited on the information it discloses. For example, an SSRF that only gives a different 
+        HTTP status for a specific type of endpoint and rejects all other endpoints (e.g. the endpoint must respond with a certain type of payload), and does not disclose any information other than the HTTP status code.
+        - Path traversal or partial path traversal that can only read access to a specific directory with limited types of files (e.g. it can only access log files, or it can only access files with a specific extension).
+        - Path traversal or partial path traversal that only discloses the existence of files or directories, and does not disclose any additional information about the file or directory.
+        - Information disclosure that only discloses whether a specific condition is true, or some 
+        id number, publicly available resources, such as source code files or documentation,
+        and does not disclose any additional information. For example, an information disclosure that only discloses whether a specific user exists, 
+        or an information disclosure that only discloses whether a specific issue id exists in the system.
+        - Issues that requires a malicious admin users to exploit during installation, configuration or other 
+        set up process. 
+        - When running CLI tools or installers, we assume the user already has control on the device. 
+        Any issues that only allows a local user to gain access of the device with running these tools or during installation 
+        is considered low severity.
+        However, do not exclude issues in installation or configuration that can be exploited by non-admin users.
+        - CSRF or XSS issues that can only be exploited during a very specific time window. For example, 
+        CSRF or XSS that can only be exploited during the installation process.
+        - Stored XSS that requires an admin user to upload malicious contents.
+        - Issue with very limited impact, such as 
+        forcing a user to log out, rate limiting, DoS issues.
+        - Misclassified issues that does not have a valid attack scenario or impact.
+        - Issues that requires knowledge of the victim's private information (e.g. secret tokens, private keys, credentials)
+        in order to exploit.
+        - Issues that requires physical access to the device or system, or sharing a device with the victim.
+        - Blind SQL injection that does not return any data, and only returns a different HTTP status code for a specific type of endpoint, and does not disclose any information other than the HTTP status code.
+
+        If you decided to reject an issue, provide the reason, and then 
+        store a low severity reason for the result with the id {{ result.result_id }}. Otherwise, 
+        you can finish the task.
+
+        DO NOT change or store anything for the current audit result.
+
+
+