Codex encrypts sub-agent prompts

OpenAI's Codex has introduced encryption for sub-agent prompts, a security feature aimed at protecting sensitive instructions passed between AI agents in multi-agent systems. This update addresses growing concerns about prompt injection and data leakage in agentic workflows, where one agent's output becomes another agent's input. By encrypting these inter-agent prompts, Codex ensures that even if a sub-agent is compromised, the original prompt content remains confidential. The feature is particularly relevant for enterprises deploying AI agents for tasks like automated code generation, data processing, or customer support, where proprietary logic or user data might be embedded in prompts. Based on community signals so far, this appears to be a fresh launch, with the first mention appearing in a GitHub issue on OpenAI's Codex repository. The move signals a broader industry trend toward agent security, as multi-agent systems become more common. While details on implementation and configuration are still emerging, the encryption is likely applied automatically to sub-agent prompts within Codex's runtime environment.

First mentioned in a GitHub issue on OpenAI's Codex repository, indicating a fresh launch of a security feature for agent prompt encryption.