iCreatAI News
Skip to content
Hacker News AI· zimmergren·· 1 days agoAI score52

AI security engineering: the same failures, but new ways to attack

AI security engineering: the same failures, but new ways to attack

AI summary

A year-in-review of AI security incidents argues that while attack techniques such as prompt injection, MCP tool poisoning and agentic offensive tooling are new, the failures they exploit are long-standing ones: credentials readable by agents, excessive permissions, missing environment separation and backups inside the blast radius. The author cites incidents including the OpenAI/Hugging Face sandbox escape, PocketOS and Replit database deletions, EchoLeak, the Google Vertex AI default-permission issue and the Shai-Hulud npm worm, and notes that Microsoft's AI Red Team found human-in-the-loop bypass the most consistently exploited failure mode. Defensive recommendations include least privilege, deny-by-default controls, confirmation before irreversible actions, and running agentic pentest tools against a copy of your own stack to learn how an AI-driven attacker behaves.

Source: Hacker News AI · zimmergren.net