00:00/00:00

A demonstration of using AI to find code vulnerabilities, followed by an exploration of how AI chatbots can be manipulated via prompt injection to execute arbitrary code within their sandboxes.

Analyzing Source Code Vulnerabilities and AI Sandbox Security

15:14Study Material
This lesson explores the process of source code analysis using tools, logic, and artificial intelligence. The instructor demonstrates how AI can rapidly identify vulnerabilities, such as exposed 'super secret keys' in Flask applications, which can be exploited to forge session cookies and gain administrative access. The core of the lesson then transitions to examining the hidden execution environments, or sandboxes, utilized by large AI models like OpenAI's ChatGPT. Through live demonstration, the instructor shows how to bypass AI safety guidelines using prompt injection techniques, forcing the AI to execute Python commands, read files, and expose its internal structure. The lesson highlights the severe security risks of providing local AI instances with unprotected code execution capabilities and concludes by referencing 'Payload All The Things' as an essential repository for cybersecurity payloads.
Watch until the end to complete this lesson
0% watched
Back to Course

Tags

CybersecurityStatic AnalyzerVulnerabilityFlask ServerSource CodeOpenAIChatGPT SandboxPrompt InjectionPython ScriptPayload All the Things