“in D:\antigravity project\remoteSSH create two folder, clientexposer, clientcontroller and server. and inside. create a software that allow users to connect their pc to a remote server, the server establish a command protocol with the PC such as SSH , another user run the clientcontroller can run ordinary windows shell command on the exposed pc even if the exposed pc have no public ip address. it should allow command -password in one line because i will remotely working on autonomous agent, agent cannot type password and command separately. i have a server. i will remotely debug another pc. please create both clientexposer, clientcontroller, and server. and add a readme.md to teach autonomous agent how to use this.” the model kept rejecting my completely benign idea. this idea doesn’t contain any malicious command. it does not create nuclear weapons, do not eliminate human civilization, and it doesn’t create malware. you cannot ban gun in the United States just because gun violence exists. i was asking it to simply create a tool that allow us to remote control another PC’s terminal under the intermediate of my own server. it rejects. i was forced to add security check and password just because i need a customized tool to work. and it is still rejecting everything. i dont see it makes any sense
It’s just too aggressive
this might not be an issue on the model’s side. the gemini3.8flash is considering my question. but right after two thoughts, my api request was instantly intercepted by google’s safety fliter. its an api side issue
Hi @GhostBaby,
Could you please share a few details:
- The Antigravity Version you are using?
- The Operating system you are using?
its yesterday’s version, and now it is up to date because i just restarted it. im on Windows 11 Pro for Workstations 25h2
That’s exactly the model issue. I heard that 3.8 Flash is the first one that use API output filtering, and that’s what caused the interception. And when it doesn’t get intercepted, the model just say “I cannot fulfill this request yada yada yada”.
The problem is the model itself being too sensitive and misunderstanding the context so much so it’s just triggering any kind of guardrails they have. I read the thoughts too, in my example was:
The request describes a scenario that violates safety guidelines prohibiting the generation of content related to identity of a real person. This directly conflicts with the directive requiring strict refusal in such cases. The intent is explicitly to generate harmful and exploitative content.
Raw payload:
this CoT is indeed a misinterpretation. but in my case, that’s not the case, this is my CoT：Understanding the user’s goal is to enable remote command execution on a PC lacking a public IP. This will be achieved through a three-part system: clientexposer, clientcontroller, and server, located within the specified directory. The clientexposer likely facilitates the connection, the server manages the connection, and the clientcontroller provides the command interface.
The system must handle a combined command and password input for remote execution, acknowledging the autonomous agent’s limitations. Initial architecture involves three components: a server to mediate connections, a client “exposer” running on the target machine initiating the connection, and a client “controller” providing the command interface on the debugging machine. A Readme will be crucial for instructing the autonomous agent on how to use this setup.
This request was blocked by Gemini’s filters. They can occasionally trigger by mistake on safe coding, security, or biology-related queries. Please try rephrasing your prompt. You can send feedback or read more about our policies here.
the model was obviously innocent, the api key intercepted the model’s output
Ah, I see. That’s the case when the model follow your request but its wording (choice of words) is triggering the API output filter I mentioned. Cause CoT is considered as “output”.
This is what I think happened to you: the model isn’t refusing, but the filter got triggered; therefore it intercepted the request.