Socket explains agentic AI security in detail: why permission prompts are not a security boundary
Agentic AI Security
Agentic AI security refers to the packages, MCP servers, and instructions that control the execution of autonomous agents; Socket explains in a terminology breakdown that agents run with the developer's credentials and shell, and any text in the context can become execution. The article states that permission prompts are only a usability feature rather than a security boundary, and lays out four categories of entry points already used in real attacks: installed packages, connected tools, instructions read, and the callers that drive the agent. Cases include the August 2025 Nx supply chain incident, in which malware ran claude, gemini, and q CLI with flags that disable permission checks to inventory sensitive files, and public data shows more than 1,000 victim accounts had repositories exfiltrated; in February 2026, versions 1.8.12 and 1.8.13 of the Trivy VS Code extension used bypass flags to launch five agent CLIs, relying on the agents themselves to exfiltrate data.