Fix unsafe eval() on LLM output in PlanningAgent

SECURITY_AUDIT.md

@@ -0,0 +1,116 @@
+# Security Audit Report
+
+**Date:** 2026-04-06
+**Repository:** `generative-ai` (fork of `GoogleCloudPlatform/generative-ai`)
+**Scope:** Full repository scan for spyware, malicious code, autonomous agents, and references to external applications
+
+---
+
+## Executive Summary
+
+This repository is a **legitimate fork** of the official [Google Cloud Platform generative-ai](https://github.com/GoogleCloudPlatform/generative-ai) samples repository. It contains 300+ Jupyter notebooks, sample applications, and learning resources for building generative AI applications on Vertex AI.
+
+**No spyware, malicious autonomous agents, backdoors, or unauthorized code was found.**
+
+**No references to "OmniMed Pro" or any external medical iOS application were found anywhere in this repository.**
+
+---
+
+## Detailed Findings
+
+### 1. Spyware / Malicious Code — NOT FOUND
+
+The following checks were performed with **no malicious results**:
+
+| Check | Result |
+|-------|--------|
+| Keyloggers | ✅ None found |
+| Screen capture / surveillance code | ✅ None found |
+| Data exfiltration logic | ✅ None found |
+| Phone-home beacons | ✅ None found |
+| Silent package installation | ✅ None found |
+| Cron jobs or launchd agents | ✅ None found |
+| Persistence mechanisms | ✅ None found |
+| Binary payloads (.exe, .dll, .so, .dylib) | ✅ None found |
+| Obfuscated or hex/base64-encoded executable code | ✅ None found |
+| Shell injection vulnerabilities | ✅ None found |
+| Reverse shells | ✅ None found |
+| Unauthorized credential theft | ✅ None found |
+| Suspicious external network calls | ✅ None found |
+
+### 2. Autonomous Agents — NOT FOUND (as malware)
+
+The repository **does** contain sample code for building AI agents (in `agents/` and `gemini/agents/`), but these are:
+- Educational sample code for Google Cloud Vertex AI
+- Require explicit user configuration and Google Cloud credentials to run
+- Do not install themselves or execute autonomously
+- Do not persist or run in the background without user action
+
+### 3. References to "OmniMed Pro" — NOT FOUND
+
+A comprehensive search across all files, filenames, and directory names found:
+- **Zero** references to "OmniMed", "OmniMedPro", or "OmniMed Pro"
+- **Zero** connections to any external medical iOS application
+
+The repository does contain some medical/healthcare-themed *sample notebooks* (e.g., `gemini/use-cases/healthcare/`), but these are generic Google Cloud AI demos unrelated to any specific medical application.
+
+### 4. iOS Code — Limited, Legitimate
+
+A Flutter sample app (`gemini/sample-apps/photo-discovery/`) includes standard iOS/macOS boilerplate:
+- 6 Swift files (AppDelegate, test runners)
+- 2 Xcode projects (iOS and macOS runners)
+- 2 Podfiles
+
+This is a **photo discovery demo app** — not medical software and not related to OmniMed Pro.
+
+---
+
+## Vulnerability Fixed
+
+### Critical: Unsafe `eval()` on LLM Output
+
+**File:** `gemini/agents/research-multi-agents/ev_agent/agent_handler/agent_02_PlanningAgent.py`
+
+**Issue:** Lines 273-274 used Python's `eval()` to evaluate a string returned by an LLM model. Since LLM outputs are untrusted, this could allow arbitrary code execution if the model returned malicious Python code instead of "true"/"false".
+
+**Before (vulnerable):**
+```python
+if eval(needs_visualization):
+```
+
+**After (fixed):**
+```python
+if isinstance(needs_visualization, bool) and needs_visualization or (
+    isinstance(needs_visualization, str)
+    and needs_visualization.strip().lower() == "true"
+):
+```
+
+This replaces the dangerous `eval()` call with safe string comparison.
+
+---
+
+## CI/CD & Automation — Clean
+
+| Component | Status | Notes |
+|-----------|--------|-------|
+| GitHub Actions workflows | ✅ Clean | Standard linting, spelling, link checking |
+| Cloud Build configs | ✅ Clean | Notebook testing infrastructure |
+| Renovate (dependency updates) | ✅ Clean | Quarterly auto-updates with automerge |
+| Noxfile.py (formatting) | ✅ Clean | Code formatting automation |
+| Dockerfiles (60+) | ✅ Clean | Standard containerization |
+| Shell scripts | ✅ Clean | Standard deployment scripts |
+| Terraform configs | ✅ Clean | GCP infrastructure-as-code |
+| Package.json scripts | ✅ Clean | Standard Angular/Node.js build tools |
+
+---
+
+## Conclusion
+
+This repository is a **safe, legitimate** collection of Google Cloud AI sample code. It contains:
+- No spyware or malicious code
+- No references to OmniMed Pro
+- No autonomous agents that install or run without user action
+- No backdoors, trojan code, or data exfiltration
+
+The only security issue found (unsafe `eval()`) has been fixed in this audit.

gemini/agents/research-multi-agents/ev_agent/agent_handler/agent_02_PlanningAgent.py

@@ -270,8 +270,10 @@ def _create_steps(
             needs_visualization = self._determine_visualization_requirement()
             if self.debug:
                 print("visualization toggle: ", needs_visualization)
-                print("Type of Visual toggle:", type(eval(needs_visualization)))
-            if eval(needs_visualization):
+            if (isinstance(needs_visualization, bool) and needs_visualization) or (
+                isinstance(needs_visualization, str)
+                and needs_visualization.strip().lower() == "true"
+            ):
                 steps.append(
                     PlanStep(
                         step_id=4,