"details": "**I. Detailed Description:** \n\nThis includes scenarios, screenshots, vulnerability reproduction methods. For account-related vulnerabilities, please provide test accounts. If the reproduction process is complex, you may record a video, upload it to Taopan, and attach the link.\n\n1. Install ms-swift\n ```\n pip install ms-swift -U\n ```\n\n2. Start web-ui\n ```\n swift web-ui --lang en\n ```\n\n3. After startup, access through browser at [http://localhost:7860/](http://localhost:7860/) to see the launched fine-tuning framework program\n\n4. Fill in necessary parameters\n In the LLM Training interface, fill in required parameters including Model id, Dataset Code. The --output_dir can be filled arbitrarily as it will be modified later through packet capture\n\n5. Click Begin to start training. Capture packets and modify the parameter corresponding to --output_dir\n\n You can see the concatenated command being executed in the terminal where web-ui was started\n\n6. Wait for the program to run (testing shows it requires at least 5 minutes), and you can observe the effect of command execution creating files\n\n**II. Vulnerability Proof (Write POC here):**\n```\n/tmp/xxx'; touch /tmp/inject_success_1; #\n```\n\n**III. Fix Solution:**\n1. The swift.ui.llm_train.llm_train.LLMTrain#train() method should not directly concatenate parameters with commands after receiving commands from the frontend\n2. The swift.ui.llm_train.llm_train.LLMTrain#train_local() method should not use os.system for execution, but should be changed to subprocess.run([cmd, arg1, arg2...]) format\n\n## Author\n\n* Discovered by: [TencentAISec](https://github.com/TencentAISec)\n* Contact: *[
[email protected]](mailto:
[email protected])*",
0 commit comments