Affected Systems
OpenAI reasoning models, specifically gpt-5.6-sol confirmed affected. GPT-6 Astra status unknown. Vulnerability exploited the compaction feature to extract encrypted reasoning output as plain text. Fixed server-side by OpenAI on October 23, 2026.
Exploitation Status
Proof-of-concept demonstrated by Xint security researchers. Public exploit code available on Xint's GitHub. No evidence of exploitation in the wild prior to fix. Method bypassed August 2026 cross-model replay protections.
Business Impact
Organizations using OpenAI reasoning models via API were at risk of proprietary reasoning process exposure, which could enable model distillation attacks by competitors. Fix was server-side; no action required by API consumers. Primary concern was intellectual property theft of model reasoning rather than customer data exposure. Researchers only extracted reasoning from their own requests, not cross-user data.
Urgency
🔵 Monitor
Recommended Actions
- Verify all OpenAI API integrations are using services post-October 23, 2026 (fix is automatic, server-side)
- Review logs for unusual compaction API calls or cross-model reasoning replay attempts between August and October 23, 2026
- If using OpenAI reasoning models for sensitive problem-solving, assess whether reasoning patterns from that period could have been exposed
- Monitor Xint GitHub repository for technical details and indicators of compromise related to GhostCompact technique
- Evaluate risk of similar compaction or context-transfer features in other LLM providers (Anthropic, Google) that may use comparable architectures