When we try to debug the OpenAI Apps Call Tool, sometimes (this error doesn’t occur every time — most tool calls succeed, and we’re not sure what the pattern is) we get this error
We checked our Cloudflare, Nginx, Service gateways but didn’t find any requests returning status 403. We’d like to ask whether this request is thrown by OpenAI, or if it’s caused by some of our own security checks?
I noticed the same inconsistency when sending tool calls via ChatGPT to my MCP endpoint.
Writing simple test data (strings like "test" and numbers like "0") worked perfectly.
But when I included a realistic price string with a currency symbol, such as "€170.00", the tool call was blocked before it reached the server, with the error:
403: Invocation is blocked on safety
Everything works smoothly once price values are symbol-free.
The error was thrown by an application policy from OpenAI. I could successfully call the endpoint manually, with the same request payload.
