Xint shows OpenAI compaction flaw can expose hidden AI reasoning

First reported by The Hacker News at · Updated · 2 sources

Security firm Xint demonstrated a way to extract hidden model reasoning through OpenAI's conversation compaction feature. The feature shortens long conversations to save space, according to The Hacker News. Researchers also moved reasoning traces between models, potentially reducing the cost of distilling a rival AI, according to SecurityWeek.

  • Reasoning is normally returned as an unreadable encrypted item, according to The Hacker News.

Covered by 2 publishers within 8 hours of the first report.

Reporting2

The Hacker News OpenAI Compaction Flaw Let Researchers Bypass Hidden Reasoning Protections · info@thehackernews.com (The Hacker News)
Cyber Security News Researchers Exploit OpenAI API Flaw to Extract Hidden GPT-5.6 Reasoning Traces · Abinaya

Related

Node.js patches 10 flaws across three supported release lines
OpenAI recording suspect Alfozan booked one-way travel to Saudi Arabia
OpenAI Wants to ‘Patch the Planet’ (With Its Own AI)
OpenAI Employee Arrested Over Hidden Recording Device

Topics OpenAI