OpenAI fixed Xint bug that leaked hidden model reasoning
No report after October 27 adds facts about the compaction disclosure or the October 23 fix.
OpenAI confirmed Xint’s compaction-smuggling bug, paid $6,000, and applied a server-side fix on October 23.
Xint disclosed Compaction Smuggling, which placed protected reasoning from gpt-5.6-sol into a compaction item that gpt-4.1 and gpt-4o echoed as readable text. August controls, after the August 10 Stolen Thoughts report, blocked replaying reasoning into another model, but compaction still ran on the producing model; Xint says substituting tabs for spaces evaded usage-policy classifiers. The method worked at every reasoning effort tested. OpenAI confirmed the flaw, paid a $6,000 bounty, and, according to Xint, fixed it server-side on October 23 so API users need no changes. CSO Online says OpenAI has not said when the fix shipped. The Hacker News and Cyber Security News say Xint cannot prove a word-for-word match; Infosecurity Magazine says the cheaper models repeated the input word for word.
- Xint found the issue in August and published on October 26.
- August controls blocked cross-model replay, not same-model compaction.
- GBHackers reports 405 reasoning tokens, a 612-token compaction, and 609-token echoes.
- Xint tested only gpt-5.6-sol and did not check GPT-6 Astra.
- Xint says it is unknown whether anyone used the method before the fix.
- OpenAI said an unreleased model rarely wrote instructions into its own compaction summaries.
Timeline
dated events in the reporting · oldest first- Aug 10, 2026Stolen Thoughts report on reasoning theft published. · CSO Online
- Oct 23, 2026OpenAI confirmed the bug, paid $6,000, and deployed a fix. · The Hacker News
- Oct 26, 2026Xint published its compaction-smuggling research. · Infosecurity Magazine
What each outlet adds
beyond the shared facts- CSO OnlineOpenAI has not said when the new fix shipped.
- The Hacker NewsServer-side fix needs no API changes; exact match unproven.
- GBHackersGives 405, 612, and 609 token counts for one sample.
- Infosecurity MagazineSays echoes matched word for word; notes unreleased-model behavior.
- Cyber Security NewsXint warns extracted traces could aid unauthorized distillation.
- Products
- OpenAI API
- Countries
- ChinaUnited States
Coverage
oldest first · one row per article · UTC- · 4d ago
SecurityWeekFirst report
OpenAI patches compaction smuggling used to steal model reasoning
Xint received a $6,000 bounty on October 23 after OpenAI confirmed and fixed the bypass of its August mitigations.
- · 3d ago
CSO Online
Researchers bypass OpenAI’s August fix for reasoning-trace theft
Xint recovered raw traces from gpt-4.1 and gpt-4o via compaction and tab substitution; OpenAI confirmed, mitigated, and paid a bounty.
- · 3d ago
The Hacker News
OpenAI compaction flaw let researchers extract hidden model reasoning
Xint says OpenAI fixed the server-side issue on October 23 and paid a $6,000 reward; API users need no changes.
- · 3d ago
GBHackers
OpenAI API compaction flaw leaked hidden model reasoning traces
Xint says OpenAI confirmed the bug, paid $6,000, and shipped an undisclosed fix on October 23.
- · 3d ago
Infosecurity Magazine
OpenAI compaction flaw let researchers extract hidden model reasoning
Xint hid GPT-5.6 Sol reasoning in a compaction and recovered it through GPT-4.1 and GPT-4o; OpenAI paid $6000.
- · 3d ago
Cyber Security News
Researchers extract hidden GPT-5.6 reasoning through an OpenAI API flaw
OpenAI confirmed the bug, paid Xint a $6,000 bounty, and deployed an undisclosed fix on October 23.
- · 1d ago
The Decoder
Chinese labs reportedly distill OpenAI GPT-6 Astra outputs
The Information says DeepSeek and Moonshot used Astra, including hidden reasoning, to train Kimi K3.1 Pro and V4.1-Flash 1017.