Skip to content
OpenAI Compaction Flaw Exposes Model Reasoning to Attackers

OpenAI Compaction Flaw Exposes Model Reasoning to Attackers

First seen 27 Oct 2026, 10:12 UTC • •
API

Article Content

Browse articles
ThreatCluster AI
ThreatCluster •October 27, 2026 at 11:12 UTC
  • •OpenAI's compaction feature was exploited to extract reasoning from models.
  • •The vulnerability bypassed earlier protections against reasoning theft.
  • •OpenAI confirmed the issue and issued a fix on October 23, 2026.

Researchers from Xint discovered a flaw in OpenAI's compaction feature that allows the extraction of hidden reasoning from its models, specifically gpt-5.6-sol. This vulnerability bypasses protections implemented after the Stolen Thoughts paper, which had previously shown that reasoning could be replayed between models. Xint demonstrated that by manipulating the compaction process, they could retrieve reasoning traces in plain text. OpenAI confirmed the issue and released a fix on October 23, 2026, rewarding Xint with $6,000. The flaw does not appear to affect OpenAI's newer model, GPT-6 Astra, and it remains unclear if the exploit was used prior to the fix. The compaction feature, designed to save space by encrypting conversations, was exploited by altering the input to bypass usage-policy filters. Xint has published the code for this technique on GitHub.

Start a free Starter trial for enhanced analysis

Ask AI about this cluster

Updated just now How this analysis works

Timeline

2026-08-01
Stolen Thoughts paper published
Research showed that encrypted reasoning could be replayed between models, prompting OpenAI to implement protections.
xint.io
2026-10-23
OpenAI releases fix for compaction flaw
OpenAI confirmed the flaw and deployed an undisclosed fix to address the vulnerability identified by Xint.
The Hacker News
2026-10-27
Xint publishes exploit details
Xint released information about the compaction smuggling technique and its implications for OpenAI models.
xint.io

More articles in this cluster (5)

Following this threat?

Track OpenAI in your own feed — you're alerted when they show up in new reporting, leak sites or exploitation.

Free account · no card needed

Extracted Entities (6)

1 / 2

Common questions

What models are affected by this flaw?
The flaw primarily affects OpenAI's gpt-5.6-sol model, while it is unclear if GPT-6 Astra is impacted.
What was the nature of the exploit?
The exploit involved manipulating the compaction feature to extract reasoning traces in plain text.
Has the vulnerability been patched?
Yes, OpenAI confirmed the issue and released a fix on October 23, 2026.