ChatGPT now on Linux, and a serious leak in encrypted reasoning traces

Two news items from 11 August that affect you as an AI-using professional in the Netherlands: OpenAI released an official desktop app for Linux, and researchers published an attack method that can extract hidden "chain-of-thought" data from large language models, including passwords and API keys.

ChatGPT desktop app for Linux: finally native

If you worked on a Linux workstation, you were until now stuck with the browser. That changed on 11 August: OpenAI launched an official ChatGPT desktop app for Linux in preview, similar to the existing versions for Windows and macOS. The app combines ChatGPT, ChatGPT Work and Codex in one native interface.

Supported distributions: Ubuntu 24.04 and 26.04 LTS, Debian 13, Fedora 43 and 44. Available as a .deb and .rpm package for x64 and ARM64.

Note: Computer Use, the feature where the model controls your screen, is not available in the Linux version outside the in-app browser. The app is closed-source and connects to OpenAI's servers. That means the privacy situation is identical to the web version: prompts and documents leave your device. For sectors with strict requirements on data location, such as healthcare, legal and financial advice, nothing changes compared with the browser.

If you want local processing, you are better off looking at open models you can host yourself, such as Meta Muse Glimmer, released on 10 August.

Source: TechCrunch, 11 Aug 2026 · Linuxiac

Reasoning traces leak: API keys found in hidden reasoning data

On 11 August, an international research team (ELLIS Institute Tübingen, Max Planck Institute, MATS Research and Snyk) published a vulnerability in the way OpenAI, Anthropic and Google return encrypted reasoning blocks via their APIs.

How does it work? Modern reasoning models, such as GPT-5.6, Claude Opus and Gemini 3, internally build a step-by-step reasoning process before giving an answer. Providers encrypt those "chain-of-thought" blocks and send them back to the client so developers can pass them on in agentic workflows. The researchers found that all models within the same model family use the same encryption key: a block originating from a powerful model can therefore be offered to a weaker variant, which can then be made to reveal the unencrypted content.

What was found? The researchers analysed 6,708 publicly available agent trajectory files and found 704 privacy-sensitive artefacts, including 62 live API keys, 33 passwords and 24 access tokens. Claude Haiku 4.5 proved the most vulnerable; a simple prefix-response technique was sufficient. In version 4.6 that specific attack vector has since been closed.

The providers concerned have received the report and state that they have neutralised the attacks.

What does this mean for GDPR compliance?

If you run an agentic workflow where personal data ends up in the context window, think of an HR tool processing CVs or a legal assistant reviewing case files, that data may in theory have ended up in the reasoning blocks and been leaked through this method. That breaches the confidentiality requirement of the GDPR.

Practical steps:

• Check whether your agent logs or trajectory files are publicly accessible; if they are, delete them immediately.
• Avoid passing encrypted reasoning blocks between different models or sessions unless it is necessary.
• Assess whether the incident triggers a reporting obligation to the Autoriteit Persoonsgegevens (report a data breach within 72 hours of unauthorised access to personal data).

For the GDPR profile of the models concerned, see our safety page.

Sources: Cybersecurity News · Simon Willison, 11 Aug 2026 · AI Governance Institute

In brief

Sources: TechCrunch, ChatGPT Linux · Cybersecurity News, reasoning leak · Simon Willison · AI Governance Institute