| Title | handrew browserpilot 0.0 Code Injection |
|---|
| Description | browserpilot is an intelligent web browsing agent controlled by natural language with LLM.
The `__step_through_instructions` method in the `gpt_selenium_agent.py` file of the BrowserPilot project is vulnerable to code injection (CWE - 94). This method uses `exec()` to run the `action` code, which may be user - inputted or LLM - generated based on user input. Although there's a `_check_danger` function to prevent dangerous code execution, it's implemented poorly. It only relies on basic string matching, making it easy for attackers to bypass.
Attackers can craft malicious Python code, like accessing sensitive files or executing arbitrary commands. They can then inject this code into the input processed by `instruction_compiler`. Once the `__step_through_instructions` method is called, the malicious code will be executed due to the ineffective security check. This vulnerability affects all versions of the relevant software, posing a serious threat to system security.
When such malicious input is processed, it may lead to the execution of arbitrary code, data leakage, or disruption of normal service operations. This vulnerability affects users who rely on the browserpilot for browser - related automation tasks.
More details: https://github.com/handrew/browserpilot/issues/21 |
|---|
| Source | ⚠️ https://github.com/handrew/browserpilot/issues/21 |
|---|
| User | ybdesire (UID 83239) |
|---|
| Submission | 04/20/2025 09:03 (1 Year ago) |
|---|
| Moderation | 05/03/2025 14:51 (13 days later) |
|---|
| Status | Duplicate |
|---|
| VulDB entry | 307195 [handrew browserpilot up to 0.2.51 gpt_selenium_agent.py GPTSeleniumAgent instructions code injection] |
|---|
| Points | 0 |
|---|