OpenAI Compaction Feature Flaw Exposes Hidden Model Reasoning to Researchers

What happened

Researchers from security firm Xint demonstrated that OpenAI’s compaction feature, intended to shrink long conversation histories, can be abused to extract a model’s hidden reasoning in plain text. By sending specially crafted inputs, the compaction routine bypasses the encryption that normally protects internal reasoning, leaking it to the caller. The vulnerability was disclosed publicly in October 2026, and OpenAI is currently investigating a fix. No CVE has been assigned yet.

Why it matters for trust and compliance

• The flaw shows why organizations must map AI‑model protection controls to a unified control framework and maintain continuous evidence of those controls for audit readiness.
• Highlights the need for continuous monitoring of AI model protection controls.
• Enables audit‑ready evidence collection for AI governance across multiple frameworks.

Who is affected

Recommended actions

1. Restrict and monitor usage of the compaction endpoint in OpenAI integrations.
2. Map AI‑model protection controls to your framework of record and collect defensible evidence.