Zero Radar
← feed
MediumVulnerabilityupdated 27 Oct 2026, 09:54 UTC

OpenAI compaction feature could expose hidden AI reasoning

A feature that shortens long chats with OpenAI's AI models could be misused to read the model's private reasoning, which is normally hidden from users. Competitors could use that to copy the model more cheaply.

Attacks reported (The Hacker News, SecurityWeek)2 independent sources
SeverityMediumVulnerability
Being exploited?ReportedBy The Hacker News, SecurityWeek, not officially confirmed yet
Fix available?UnclearNo fixed version published in the sources yet
How sure is this?2 independent sourcesSeveral independent publishers agree

How the attack works simplified

  1. 1Attacker uses compactionShrinks a long conversation on purpose
  2. 2Reasoning exposedHidden reasoning emerges in plain text
  3. 3Traces copiedReasoning captured and moved between models
  4. 4Rival model distilledCompetitor trains a cheaper copy
  5. ResultHidden AI reasoning is copied

What happened

Researchers at Xint showed that OpenAI's conversation compaction feature can be abused to get encrypted reasoning items out of reasoning models in readable plain text, bypassing the protection that keeps them unreadable to users.

Why it is dangerous: OpenAI's hidden reasoning is valuable intellectual property, and exposing it could let rivals, including reportedly Chinese developers, distill cheaper competing models. Anyone relying on the confidentiality of these reasoning traces is at risk.

Are you affected?

You are affected if you use OpenAI reasoning models with the compaction feature enabled and depend on their hidden reasoning staying private.

OpenAI reasoning models using the compaction feature (exact model and version list unknown).

What to do

  1. Monitor OpenAI's official channels for an announced fix or mitigation to the compaction feature
  2. Limit sensitive prompts sent to reasoning models until a fix is confirmed
  3. Review whether your usage depends on keeping reasoning traces confidential

Timeline

  1. 26 Oct 2026, 14:30 UTC
    Publicly disclosed
  2. 26 Oct 2026, 14:30 UTC
    Picked up by Zero Radar
  3. 27 Oct 2026, 08:22 UTC
    Reported by The Hacker News · The Hacker News
  4. 27 Oct 2026, 09:54 UTC
    SecurityWeek reported on it · SecurityWeek

Sources 2 reports, oldest first

Security press established security newsrooms

Technical details

ID
ZR-ea5496fe · ZR-38ba7bc2
Disclosed
26 Oct 2026, 14:30 UTC
On Zero Radar
26 Oct 2026, 14:30 UTC (0 min after disclosure)
Last change
27 Oct 2026, 09:54 UTC

The plain-language parts and the attack diagram were written by an AI model (anthropic/claude-haiku-5.5) from the sources above. Facts like versions, exploitation and dates come from the sources themselves. How Zero Radar works

OpenAI compaction feature could expose hidden AI reasoning · Zero Radar