Responsible Disclosure: My Experience Reporting to Anthropic

When you work daily with AI systems, you start looking at every interface with a pentester's eyes. That is how I discovered a vulnerability in Anthropic's Claude Code. I was not hunting for bugs at that specific moment - I was using the tool for work. But years of offensive security experience teach you to notice anomalous behavior, even when you are not actively testing.

The vulnerability involved a pattern I would later find in many other projects: how software handles credentials during HTTP redirects. I won't go into technical details out of respect for the disclosure process, but the bug class is significant because it affects functionality that developers often consider "secure by default."

The disclosure process

Reporting to Anthropic was a positive experience. The security team responded quickly, took the report seriously, and worked on the fix with transparency. This is the ideal model of responsible disclosure: the researcher finds and reports, the vendor acknowledges and fixes, the ecosystem becomes more secure.

What impressed me was the maturity of their process. No minimization, no "that's not a real bug." A clear acknowledgment, a fix timeline, and constant communication. It is exactly how it should work, and unfortunately not all vendors behave this way.

This experience triggered my broader vulnerability research campaign. If a bug pattern exists in a high-profile project like Claude Code, it probably exists elsewhere too. And indeed it did - but that is another story.