OpenAI just stripped developers of audit access. 17 days before the EU AI Act deadline.

OpenAI shipped Codex CLI 0.144.4 on July 14.

Since then, developers running Sol or Terra models can no longer read what their parent agents told subagents to do.

The instructions are encrypted. Only OpenAI's servers hold the key.

This is not a bug. It's a architecture change merged on June 5 in PR #26210.

Here's the problem: EU AI Act Article 12 takes effect August 2, 2026. That's 17 days.

Article 12 requires high-risk AI systems to maintain logs detailed enough to reconstruct what the system did and why.

When a Codex pipeline runs on Sol or Terra, delegation text is now ciphertext in local records. You can see that a handoff happened. You cannot read what was delegated.

For enterprises using Codex in critical infrastructure, financial services, or employment decisions — you just lost the audit trail Article 12 demands.

OpenAI's response: silence. No public rationale. No timeline. No clarification on whether their enterprise Compliance Platform retains plaintext copies.

GitHub Issue #28058 proposed a fix: keep the encrypted delivery channel, add a parallel plaintext audit field for local inspection.

OpenAI hasn't committed to implementing it.

The asymmetry is clear. OpenAI can read your delegation instructions. You cannot.

Audit your Codex pipelines today. If you're running Sol or Terra in any context that could trigger high-risk classification under Annex III, you have a compliance gap that opens in 17 days.

SOURCE: https://www.techtimes.com/articles/320784/20260716/openai-codex-encrypts-agent-instructions-stripping-developers-audit-access.htm
VERIFIED: TechTimes (July 16, 2026), GitHub PR #26210, GitHub Issue #28058, EU AI Act Article 12
SIGNAL: OpenAI just made enterprise AI governance harder 17 days before the EU's highest-stakes compliance deadline. Every enterprise running Codex multi-agent pipelines needs to assess their audit posture now.