Threat video1:00
OpenAI API Reasoning Leak via Compaction: Detection & IOCs
OpenAI's compaction endpoint returned encrypted GPT-5.6 reasoning as readable text: 9 detection rules, 9 IOCs and a free Splunk SPL rule.
Read the full threat intel for OpenAI API Reasoning Leak via Compaction: Detection & IOCs →
Chapters
- 0:00The hookCompaction smuggling extracts about 609 tokens.
- 0:04The storyPrior attacks were mitigated in August.
- 0:11Attack pathThey harvest encrypted reasoning from gpt-5.6.
- 0:27AttributionNo threat actor: this was defensive research.
- 0:32Indicators of compromise5 indicators to hunt: 2 domains, 2 URLs and 1 file name.
- 0:38MITRE ATT&CK techniques3 ATT&CK techniques, from initial access to collection.
- 0:43Remediation3 steps now: No developer action required.
- 0:50Detection rules9 detection rules, in SPL, KQL and Sigma.
Transcript
0:00Compaction smuggling extracts about 609 tokens.
0:05Prior attacks were mitigated in August.
0:07Researchers found a bypass in October.
0:11They harvest encrypted reasoning from gpt-5.6.
0:15Researchers reproduced compaction with preserved reasoning.
0:19Swap spaces for tabs, evade filters.
0:22Weaker models echo the plaintext back.
0:27No threat actor: this was defensive research.
0:325 indicators to hunt: 2 domains, 2 URLs and 1 file name.
0:383 ATT&CK techniques, from initial access to collection.
0:433 steps now: No developer action required.
0:519 detection rules, in SPL, KQL and Sigma.
0:56Threadlinqs. The unified security engineering platform.