Codex starts encrypting sub-agent prompts (github.com)

Article summary

Codex has started encrypting sub-agent prompts, which removes the human-readable task audit trail. This change affects versions that include the update and enable MultiAgentV2. The goal is to preserve encrypted delivery while restoring a local audit trail. A proposed implementation keeps the existing encrypted message field and adds a non-encrypted plaintext companion to each v2 communication tool.

Main themes

• Codex encryption
• MultiAgentV2
• Audit trail
• Encrypted delivery
• Local auditability
• AI model integration

What commenters say

• The encryption change may not significantly impact users who utilize alternative harnesses or custom integrations with the Codex app-server.
• Some users are concerned that this change could be a step towards restricting the use of GPT subscriptions in alternative harnesses.
• The encryption is seen as a way to prevent collection of queries from users and resellers, and to stop the storage and resale of data.
• Others believe that the change is not a significant issue and that the benefits of encrypted delivery outweigh the loss of human-readable audit trails.
• There are differing opinions on the potential impact of this change on the use of Codex and other AI models, with some seeing it as a minor issue and others as a major concern.
• Some commenters are skeptical about the motivations behind the encryption change, while others see it as a necessary step for security and privacy.
• The discussion also touches on the topic of homomorphic encryption and its potential applications in AI model inference.
• The change is seen as a way to prevent black market resellers from storing and selling data, and to protect user privacy.