A new CVE against Cursor, a LiteLLM supply-chain backdoor, and a study showing image-only injection hits 76% success rate — all in the last 72 hours. Here's what each one means for agents in production.
*Three prompt injection stories from this week that your guardrail probably missed
*
Three things happened in the last 72 hours that, taken together, describe exactly where the agentic AI attack surface is heading. None of them involved an exotic zero-day. All of them exploited the same structural gap: agents that trust what they read.
Here they are, with what each one means for engineers shipping agents right now.
