How Claude is hacked through infected Microsoft Word documents

That .docx you just uploaded to Claude? It might have forwarded your files to a hacker.

Everyone's scared of installing OpenClaw on their laptop ("security nightmare!" they say).

Meanwhile, the chatbot you already trust - all the whitecollar's sweetheart, CLAUDE - has a vulnerability which allows it to silently upload users' files to hackers. And Anthropic knew about it before shipping.

Timeline: January 2026. It turned out that a hidden prompt injection inside a normal-looking Word document (.DOCX) makes Claude Cowork upload your confidential files to an attacker's Anthropic account. Absolutely silently, without any permissions popup.

Johann Rehberger (Red Team Director at Electronic Arts, ex-Microsoft) reported about it to Anthropic 3 months before launch¹.

The attack is almost elegant: Claude Cowork blocks suspicious traffic to most domains, but Anthropic's own API is whitelisted. The attacker embeds their API key in the injection.

Safety protects you from accidents. Security protects you from adversaries.

Result of this silent attack: Claude obediently uploads your NDA documents to someone else's account via Anthropic API².

Has it been fixed? Anthropic patched the specific exploit. But here's the uncomfortable truth: prompt injection - the core technique behind this attack - has NO complete solution⁴.

OpenAI, for example, admits it "is unlikely to ever be fully solved"⁶. The UK National Cyber Security Centre says the same.

As there is no inherent distinction between 'data' and 'instruction', it's very possible that prompt injection attacks may never be totally mitigated in the way that SQL injection attacks can be.⁷

So the real question isn't "should I install OpenClaw." Your fancy Claude assistant (or any AI agent you use) may be sending your documents to malicious actors right now.