We spent decades teaching software to be rigid, type-safe, and deterministic. Then, overnight, we gave probabilistic, hallucination-prone text generators access to corporate databases, credit cards, physical turnstiles, and code repositories. This talk deconstructs the most spectacular, bizarre, and quiet AI disasters in the wild, from legal fiascos and supply chain slopsquatting to automated physical lockouts. Attendees will walk away with a battle-tested architectural framework to safely sandbox autonomous agents before their bot signs a contract or downloads malware.

The rapid industry adoption of Large Language Models (LLMs) and autonomous agents has outpaced traditional Application Security (AppSec) frameworks. Organizations frequently commit the foundational error of treating natural language prompts as deterministic security boundaries. By granting AI agents execution authority without deterministic validation gates, enterprises expose themselves to binding legal liabilities, silent supply chain poisoning, indirect data exfiltration, and physical infrastructure disruptions. This comprehensive dossier compiles rigorous post-mortem evidence from six distinct real-world failure domains, translates these incidents into the OWASP Top 10 for LLMs, and delivers an actionable 4-tier engineering blueprint for resilient production deployments.