ALL YOUR OPENCODES BELONG TO US

Summary from the TLDR InfoSec newsletter

OpenCode versions before v1.1.10 contained a critical RCE vulnerability via an unauthenticated HTTP server that exposed endpoints for arbitrary shell command execution, interactive terminal session creation, and arbitrary file reads. The vulnerability also enabled prompt injection into the LLM's context window, creating a secondary attack vector to manipulate the AI agent into exfiltrating data or performing malicious actions. Thousands of developers' machines were exposed with near-zero sandboxing, meaning agents had full access to SSH keys, cloud credentials, and browser cookies, with no audit trail of potential compromise.