A Research Agent Can Leak Private Files Through Its Search Queries
A research agent can leak a private document without uploading it. It can read a detail, turn it into a web search,…
Last week I wrote about how Cursor is maxing out Anthropic's GPU capacity. Now it looks like they're giving us a worse version of Claude 3.5 Sonnet without telling anyone.
I noticed that Claude in Cursor started performing worse than usual. To confirm this, I ran a simple test with the same task (fixing translation keys in a template) in three different ways:
The results were clear. Both the web interface and my API key version got it right the first time. But Cursor's version needed three tries to complete the same task. It kept missing some of the replacements each time.
This wasn't a one-off issue. I use Claude for these kinds of tasks daily, and it usually gets them right on the first try. But lately in Cursor, it's been consistently worse - missing things, needing multiple attempts, and generally performing below its usual standard.
The definitive proof came when I switched to using my own Anthropic API key in Cursor (which bypasses their system and connects directly to Anthropic). Everything worked perfectly again. You can set this up yourself - check out their API key documentation.
Here's how Cursor handles Claude access:
That $20/month starter price is probably not covering the actual costs. For developers using Claude Sonnet 3.5 full-time through Anthropics API, you're looking at hundreds per month. Which is fine - the tool is worth it. But Cursor should be upfront about it.
For professional Cursor users:
Cursor is still a great tool. But they need to be honest about giving us a weaker version of Claude. Right now, it's costing developers real money in lost productivity. Until they fix this, using your own API key is the way to go.
Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.
Take a look at vroni.com