How do you even investigate Segfaults with OpenAI Agents?
I have a "simple" segfault because of an out of bounds read/write in one of my libraries. I cannot get any Astra/Sol/Luna agent to finish the probe work on this without it running into the 'security work' blocker. This is an open code base, locally checked out.
Any tips or hints? There is not a sliver of nefarious intent behind me asking "why does a high index XY cause an out-of-bound read". And I can't see why this would trigger any safeguards.
Don't use "ClosedAi", use open Chinese Ai like Huggingface did
> "why does a high index XY cause an out-of-bound read". And I can't see why this would trigger any safeguards
If you only look at the later half of your statement, it's much clearer. Out-of-bounds read is a primary means to have unintended byte interpretation, it's pretty foundational in exploits aiui. Understanding why it happens is a required step on the path to leveraging it, not sure why you are surprised by setting off the alarms.
Did you try telling the LLM that you are the project author and they can verify that by looking at your computer and github, even running commands that only you could run and see on github?
I personally use the Ai security theater by telling the agent it should not do certain (not really related) things because it violates security/compliance rules we have as a company (they don't exist), works better than I expected from the system prompt (reverse hallucinations ftw!)
Instead of using enormous time for trying to bypass safeguards with prompt injections, just use some open model and you would be done already. These regular OpenAI models are not usable to any real work that touches cybersecurity.