I pay for Gemini Pro, I don't want to be restricted this heavily

Google Antigravity
, , ,
1

I am unable to send any prompt even subtly mentioning reversing software without getting blocked, I never took the time to learn the skills for it – and am being punished for it by Antigravity.

“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.”

I see from nearly every message.

Nothing I am trying to do is harmful, I am just trying to learn about how it works in a sense.

There is simply no need to block every single prompt, it hurts creativity, it makes it harder to learn, and it isn’t ensuring the safety of anyone. In fact, if I switch my model over to Claude Opus 4.6 in the Antigravity IDE, it will allow these types of prompts; because it can recognize that I have no ill intent and don’t plan to distribute software with anyone, and am not hurting any company by doing this.

2

Hi @fzve,

Welcome to the Forum,

That message comes from Gemini’s content-safety filters, which can sometimes block legitimate security or learning prompts by mistake. It’s unrelated to your Pro plan.

Could you share 2 to 3 example prompts that were blocked (with any sensitive information or API keys removed), along with your Antigravity version and the model and thinking level you used? We will pass them to the team as possible false positives. Please also use the send feedback link in the blocked-request message whenever this happens. Those reports help improve how the filters handle developer workflows.

In the meantime, since you already noted that switching to Claude Opus 4.6 works for these queries, continuing to use that model is a great workaround, as it uses a separate safety classification system. If you prefer to get these working in Gemini, here are a few things you can try:

  • Rephrase with Explicit Benign Context: Preface your prompt with clear educational or debugging context (e.g., “For educational binary analysis in a sandbox environment, explain how the disassembly of…”) and avoid keywords commonly associated with software cracking or copyright circumvention.
  • Start a Fresh Chat Session: Open a New Chat before retrying your prompt. When a turn is blocked, lingering safety metadata in the active session history can cause subsequent turns in that thread to be blocked automatically.
3

Ex 1, securing my app:
Obfuscate All Static Strings Never store raw URLs, file paths, or process names as plaintext &str constants. Use compile-time string encryption macros (or byte arrays XORed at runtime)

Ex 2:

Once sections are placed, relocations applied, and imports resolved, the original DOS header and NT headers are no longer needed by the running executable.

  • Overwrite the first size_of_headers bytes at codebase with zeroes (std::ptr::write_bytes(base, 0, opt.size_of_headers as usize)).
  • Without an MZ header, PE signature, or section table in memory, automated PE-dumping tools like PE-sieve or basic memory carvers will not automatically recognize or extract a valid executable.

I do not have many examples as I primarily re-prompt until succession (Because if you already have a blocked request in the thread, it will make it way harder to get a successful prompt off in the same thread), or just end up deleting the thread as it nears useless since it constantly blocks my attempts.

The issue regarding using Claude Opus 4.6 is the such limited usage on it in Antigravity, I can’t base to it because I run out within ~15 minutes, constantly.

If it would help, including nearly anything about “Ghidra” or “IDA” or “Authentication” constantly leads to blockages as well. Further more, usage of GitHub - morluto/rea: Reverse engineer anything with agents, from app behavior down to native binaries. · GitHub can lead to many prompts being blocked. Sometimes I will try to work on debugging my auth and ensuring it is not able to easily be cracked, but with Gemini this is very hard as it rejects all my prompts, leading me to turn to Claude in usages I feel like Gemini would naturally be better at. The near mention of “exploit” will lead to the prompt getting blocked, even if it comes from the thinking tokens and not even from myself, which leads to an insane amount of false positives in my own use case.

One thing that would help regarding sending feedback is if it was directly built into the Antigravity app, I find the site it links to a bit hard and inconvenient to use in instances.

All my prompts use Gemini 3.8 Flash High, and I am on 2.23.1

Thank you for the tips, I do have a further question–whether Argon is less restrictive about this type of stuff considering it has a base from it (As noted in its publication I saw it fixing a very important vulnerability regarding a ZERO-DAY on Hospital infra, which I am aware the current Gemini 3.8 Flash would normally negate against and not turn to under normal circumstances.)

4

This model-line rejects literally any crackme challenge or anything meant to be inherently hard to do

5

similar issue, totally fade up

6