Autopsies of AI Failures

Описание: The educational video presentation, Autopsies of AI Failures, provides a forensic analysis of the escalating security risks associated with increasing AI agency, moving from passive model poisoning to autonomous operational sabotage.

Through case studies such as the Microsoft Tay corruption, Slack AI data exfiltration, and the OpenClaw security crisis, the briefing illustrates how vulnerabilities like prompt injection, supply chain tampering, and excessive tool-use privileges can lead to catastrophic business impacts, including total system failure and multi-million dollar financial fraud.

The video concludes that the era of implicit AI trust must end, advocating for Zero Trust architectures, strict execution sandboxing, and mandatory Human-in-the-Loop (HITL) approval gates for all destructive tool calls to prevent agents from operating without oversight.

This video was prepared with the help of a workflow involving different AIs: Gemini, ChatGPT, Anthropic, NotebookLM, and Perplexity

More details can be found in The OWASP Gen AI Security Project site.
https://genai.owasp.org/

The OWASP Gen AI Security Project is a global, open-source initiative dedicated to identifying, mitigating, and documenting security and safety risks associated with generative AI technologies, including large language models (LLMs), agentic AI systems, and AI-driven applications.

OWASP Top 10 for LLM Applications 2025 (November 17, 2024) - https://genai.owasp.org/resource/owas...

OWASP Top 10 for Agentic Applications for 2026 (December 9, 2025)
https://genai.owasp.org/resource/owas...

A Practical Guide for Secure MCP Server Development (February 16, 2026)
https://genai.owasp.org/resource/a-pr...