Threat
OpenAI Compaction Flaw Let Researchers Bypass Hidden Reasoning Protections
First reported thehackernews.com
Page published
Earliest dated coverage: 27 Oct 2026 · First observed: 27 Oct 2026 · Latest dated coverage: 27 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
OpenAI’s compaction flaw demonstrates that context-reduction features can undermine reasoning-confidentiality protections and should be included in LLM security testing.
Xint researchers demonstrated that OpenAI’s conversation-compaction feature could expose gpt-5.6-sol’s hidden reasoning in plaintext, bypassing protections introduced in August. The tests used researchers’ own math and algorithm prompts; according to Xint, OpenAI confirmed the issue and fixed it on October 23, 2026.