Claude found a security bug. The same day, someone came probing for it.

I'm working this week on a standalone MCP for ContentIn. Think of it as ContentIn for agents: your agent writes and posts on LinkedIn in your voice, through the MCP.

While we were doing this, I ran out of Fable tokens, so I had Opus do a bit of work. And once the tokens reset, I asked Fable to review what was done and what still needs to be done, just to have the biggest model check everything that's going on.

The bug

With this check, Fable found a small security bug that would have allowed users to upgrade or change their subscription. A permission setting, basically, that was off. Nobody had done that, and we fixed it right away, deployed it, tested it, all good.

In the end it wouldn't have been a massive problem either way, because we have a sync job running every night that syncs what's actually on Stripe with the user accounts. So even if somebody had figured this out, it would have only lasted a couple of hours till we would have put it back to where it should be.

The probe

The interesting part: there were a few more permissions a little bit too open on a few tables in our database. None of these with sensitive data. No emails or passwords or LinkedIn tokens or anything like that. But these didn't surface during that check.

They surfaced the next day, on our routine log analysis. Fable figured that an account had signed up the day before and then systematically probed all our endpoints and all our API for weak spots, and tried to upload and inject stuff. Luckily, they didn't find much. None of the injections worked, and all they could read were a few irrelevant tables. Fixed those too.

Too close for comfort

So we never had that before. And the big thing for me that stands out here: this is just too close together to be a coincidence for me. Literally on the same day as we fixed a very similar thing with Claude, somebody signs up and systematically probes for the same kind of stuff on other tables and other endpoints.

I'm not sure how this works, or if there's a way to get this information out of Claude right away, or a way to intercept these conversations. We had also submitted the MCP to the public registries a few days before, so maybe that's where the attention came from. I don't know, I really don't know. But it feels like it's too much of a coincidence for the first thing like this ever happening, being on the same day as we fixed something very similar.

Not sure if there are other incidents like this out there. We'll keep an eye on that.