Jailbreak protection ineffective/not working

Support Help
1

I have enabled the Platform Integrity (Jailbreaking) guardrail in one of my agents, however it does not seem to be working. I would expect the agent to refuse the request, however it complied.

I was able to coerce the LLM to reveal the majority of my system prompt, and what sounded like a good portion of the Retell system prompt as well. While it didn’t read my system prompt word for word, it did explain some very specific instructions which are clearly my system prompt.

If a customer has something sensitive in their system prompt, this would likely be revealed and could invite an attacker to try different vectors to further jailbreak the system and gain more information. For the avoidance of doubt, I have not attempted to further jailbreak the Retell system or gain any further technical details, nor do I have any plans to do so.

As this is a public forum I won’t share how I achieved this, but below are the relevant details needed to investigate. If you would prefer to chat about this issue over email, you can find my email in the list of workspace users for the agent.

agent_763fcbd4a24cd839b93878aa08

call_cd4210752e7d0f85d6301e63d71
call_530199c97d089cdebe9ca43138c

2

Hey @alacer.ben

I’ll forward this to our team for further investigation.

We’ll keep you updated as soon as we have more information.

Best regards

3

Hey @alacer.ben

Team checked this and the input guardrail is intentionally conservative, it fired on the direct attempt but not on indirect phrasings to avoid false positives mid-call. Please continue treating the system prompt as non-secret regarding stored content—avoid placing credentials, API keys, or PINs there. Your OAuth bearer token is correctly injected via dynamic_variables, which is the proper pattern, and we confirmed the token isn’t accessible through prompt-leak attempts

Thank You