“But marinade” and leaked passwords are what researchers found in ChatGPT’s hidden reasoning

What happened

Security researchers uncovered a serious vulnerability in the APIs of major AI providers including OpenAI, Anthropic, and Google. This flaw allows attackers to extract encrypted reasoning traces—the internal step-by-step logic models use to generate responses—and move them between different AI models. A public scan revealed dozens of passwords and API keys floating in these traces. Moreover, the reasoning summaries seen by users often hide what the AI models are truly doing behind the scenes.

The risk

These leaked reasoning traces expose sensitive credentials and internal logic that should remain private. Attackers with access to the API could intercept passwords and keys, creating clear security risks for businesses relying on these AI services. The ability to transfer traces between models also threatens the integrity and confidentiality of AI workflows, potentially allowing advanced exploitation across platforms.

Why it matters

This vulnerability pressures AI providers to tighten data isolation and encryption in their APIs to prevent accidental credential leaks. It lowers trust in model reasoning transparency by showing that the simple summaries users see are often sanitized or incomplete. Operators who handle sensitive data through AI models face increased risk of credential theft or data exposure, raising costs around security audits and mitigation. This weakens one of the core promises of large language models: secure, private handling of user data and reasoning.

Who should pay attention

Businesses integrating AI models into workflows with confidential or regulated data should urgently reassess their security posture. Founders and operators relying on OpenAI, Anthropic, or Google APIs must demand better protection of internal traces from providers. Security teams should monitor for suspicious activity linked to leaked keys or passwords. AI engineers building on these APIs should plan for enhanced encryption and consider isolating sensitive reasoning steps.

What to watch next

Watch for security patches and API updates from these AI vendors addressing trace data leakage. Follow how providers improve transparency and visibility into model reasoning without exposing sensitive details. Regulators and enterprise buyers may increase scrutiny over data protection guarantees from AI suppliers. Vendors who swiftly fix these issues will strengthen trust and maintain business relationships at a time when breaches can rapidly erode confidence in AI adoption.

AI Quick Briefs Editorial Desk