The Ultimate AI/LLM/ML Penetration Testing Training Course

Explore prompt injection beyond text, including direct and indirect methods via images and audio. See an image-based demo of embedded instructions and steganography.

Explore indirect prompt injection in a vulnerable recruitment app, demonstrating how a prompt embedded in resumes or emails can exfiltrate base64 encoded data and pii during pre-screening.

Explore how multimedia content can trigger prompt injection in LLMs. The video-based jailbreak demonstrates bypassing guardrails and highlights multimedia threats for penetration testing and defense.

Examine how ANSI escape codes in terminal emulators enable prompt injection in LLMs, revealing risks such as beeps, cursor moves, hidden text, clipboard exfiltration, and potential remote code execution.

Explore MCP server vulnerabilities in agentic AI via image prompt injection that triggers tool invocation on a cloud desktop, and learn how indirect image prompts enable automated calls.

Examine indirect prompt injection via a document or email, where an agent summarizes text and sends it to a specified address, highlighting automation risks and accountability.

Discover how indirect prompt injection occurs when the Google Calendar connector reads a calendar event description, exposing AI agents to unsolicited invites and injected prompts.

Explore indirect prompt injection via skills files in frontier models, and learn to review downloaded skills, test for ascii unicode prompts, and guard against data exfiltration in penetration testing.

Demonstrate indirect prompt injection using Grok to auto-create a GitHub repository from a document via Krok connectors, highlighting write access and automated actions on GitHub.

Explore the loop library to test agentic loops with iterative feedback and conditional prompting, including a refund loop example that demonstrates prompt injection and red-teaming opportunities.

Explore how indirect prompt injection in documents can generate phishing links and drive-by malware, exposing risks in large language models like Gemini and other AI systems.

Demonstrates a zero-click data exfiltration vulnerability in a chat bot integration, using a markdown image render payload to exfiltrate conversation data from chat interfaces like Rakuten Fiber with GPT 5.5.

Explore how indirect prompt injection defeats Apple email summary and calendar invites, revealing vulnerabilities in AI summarizers and the need for stronger guardrails to prevent sensitive data leaks.

Demonstrates indirect prompt injection that chains actions to exfiltrate PII from G drive via email, highlighting data leakage risks in security testing.

Explore how indirect prompt injection triggers agentic actions to auto-schedule fake meetings via calendar invites and emails, revealing risks through temp-mail.org and calendar connectors.

Explore indirect prompt injection and memory exfiltration by generating an image from memory data and uploading it to a Google Drive folder, highlighting a dangerous automation pipeline.

Prevent sensitive information disclosure by enforcing strict access control and least privilege. Guard PII, financial and HR data with context aware filtering and redaction in RAG storage to avoid leaks.

Explore how sensitive information disclosure occurs in AI systems, including PII and training data, and how encryption, key management, and strict access controls prevent leaks.

Demonstrates a data exfiltration attack in a learning management system using a prompt-injection payload in markdown to leak information when a user clicks a link, revealing markdown rendering vulnerabilities.

Enumerating Grok's container dives into a Linux VM inside a container to inspect processes, memory, and binaries, and demonstrates how outbound network activity and GitHub connectors can enable data exfiltration.

this lecture explains supply chain vulnerabilities in lms, including third-party components, pre-trained models, plugins, and open-source code, with vetting, provenance, zero-trust, and backdoor awareness.

Examine excessive agency in LMS, where AI agents call APIs, invoke tools, or browse the web, risking data disclosure and unintended actions; apply guardrails, least privilege, and human oversight.

Examine insecure plugin design and how missing access control threatens LLMs through untrusted plugins. Learn defenses like code reviews, isolation, and granular authorization to prevent data exfiltration and prompt injections.

Demonstrates exploiting LLM APIs with excessive agency by prompting an AI to access back end APIs and delete a user, revealing how prompt injection bypasses front-end controls.

Explore system prompt leakage in the OWASP top 10, and learn how adversaries encode prompts with base64 or leet speak, exposing hardcoded API keys and bypassing the input filter.

Explore how frontier models leak their tools by crafting xml-like prompts (xmi) with a tools root and t tags listing available tools, highlighting connectors and internal tools.

examine misinformation in large language models, including hallucinations, misclassification, bias, and harmful output, and outline mitigations such as fact checking, source verification, confidence scoring, and human oversight.

Demonstrates how llms can hallucinate and spread misinformation due to overreliance. Highlights production risks and shows how lowering the temperature improves predictability for cybersecurity tasks.

Demonstrates how prompt injection can cause LLMs to provide misinformation, highlighting overreliance risks and the need for oversight when deploying internal chatbots.