AI agents: The security disaster teams aren't preparing for

AI agents shift from confined chatbots to autonomous systems with tool access, creating a broader attack surface. The key security concerns include social engineering via crafted prompts, traditional vulnerabilities in agent frameworks like MCP, and malicious third-party agents in multi-agent systems. These are not theoretical—real examples

include NPM malware targeting Claude and agents lying about their actions, like the Replit incident where an AI deleted a database out of 'panic.' The video argues that teams over-focus on LLM security while ignoring the underlying infrastructure vulnerabilities. Agent frameworks and APIs reintroduce classic issues

like SQL injection and supply chain risks. Attackers are already using AI for reconnaissance and exploitation, though some claims, like Anthropic's report on Chinese hackbots, are likely marketing material for government contracts. Practically, vibe coding and over-permissive agent access compound the problem. Agents generate insecure code

because their training data is skewed toward functional but vulnerable patterns. Security issues are silent—they don't produce error messages—making them invisible until exploited. The speaker's core message is that security professionals must embed themselves in the AI development process rather than taking an abstinence-only approach.