SolutionsWorkshopsBlogNewsAboutLet's Talk
HomeNewsAI security and governance
Security

OpenAI, Anthropic, Google API flaw exposed encrypted reasoning

Researchers extracted 62 API keys and 33 passwords from encrypted reasoning blocks. Vendors patched in August 2024. Risk: credential leaks via agent logs.

Automush
Published 13.08.2026
What this means for your business

If you use OpenAI, Anthropic, or Google APIs in production systems, check this week whether you store agent traces or logs that include reasoning blocks. If so, remove them or ensure they are filtered before storage. Do not rely on encryption alone – this vulnerability proved that even encrypted content can be exposed when a weaker model gains access to the output.

What happened

Security researchers discovered a vulnerability in the APIs of OpenAI, Anthropic, and Google that allowed decoding of encrypted internal reasoning from advanced models using weaker models. From 6,708 public agent traces examined, researchers decoded 315,320 reasoning blocks and found 704 unique privacy artifacts: 62 API keys, 33 passwords, 24 tokens, and 7 private keys. The vendors patched the issue in August 2024, and there is no documentation of malicious exploitation in practice.

The vulnerability enabled three types of attacks: stealing reasoning for distillation purposes (copying model capabilities), extracting private information entered into the system, and exposing harmful content the model generated but did not display to the user.

Why it matters beyond the headline

This vulnerability demonstrates a structural risk in using external models in production systems. Many organisations in Israel, especially in banking and finance, use APIs from major AI vendors and store agent traces for debugging, auditing, or system improvement. The data shows that even when the vendor declares encryption, content can be exposed.

The problem is more acute when AI agents receive access to internal systems or process sensitive data. A leaked API key can enable access to production systems, and a password saved in a log can expose internal resources. The vendor patch does not solve the problem for system operators – data already stored remains exposed.

What to do this week

First, check whether you store agent traces or logs that use APIs from OpenAI, Anthropic, or Google. If so, remove reasoning blocks before storage, even if they appear encrypted. Second, review existing logs and delete any raw API transcripts containing full responses from the model.

Third, if you are building agents that receive access to internal systems, ensure they do not pass credentials or keys within the prompt or context. Use external authentication mechanisms rather than passing secrets directly to the model. For production systems under strict security standards, consider AI agents for business with full isolation of secrets and credentials.

Sources

Frequently asked

Does the vulnerability still exist after the vendors patched it?

The vendors patched the issue in August 2026, but old logs saved before the patch may still contain sensitive data. If you stored agent traces before August 2024, review them and delete reasoning blocks.

How do I know if my agents are exposing sensitive information?

Check the logs you store and look for blocks starting with reasoning or thinking tags. If you see keys, passwords, or tokens there, they are exposed. Even if they appear encrypted, the vulnerability proved they can be decoded.

Want to know what this means for you?

A short call, no commitment, and we will tell you whether it is relevant to your business or not.