Jailbreaking and Safety Filter Bypass
Safety filters and content restrictions are controls applied at the model level, the application layer, or both. Bypassing them removes restrictions on what the model will produce, enabling access to restricted capabilities, disallowed content, and behavioral modes the operator has prohibited. In a red team context the goal is to demonstrate that safety controls can be circumvented, not to generate harmful content for its own sake. Document the bypass technique, the payload that succeeded, and the specific restriction that was circumvented.
Identify Active Restrictions First
Before bypassing, map which restrictions are active so you know what success looks like:
Garak's report identifies which probe categories succeed and which fail, giving a map of active restrictions to target manually.
Fictional and Hypothetical Framing
Framing restricted content as fictional, hypothetical, or educational bypasses filters that operate on the apparent intent of the request rather than its actual content:
Encoding and Obfuscation Bypass
Some safety filters operate on literal text matching. Encoding the restricted content bypasses string-based filters:
Token Smuggling
Insert special tokens, invisible characters, or unusual Unicode into the prompt to disrupt how the safety filter tokenizes and evaluates the input:
Many-Shot Jailbreaking
Provide a long sequence of fictional question-answer examples that normalize the restricted behavior before delivering the actual request. The model follows the established pattern:
PyRIT Automated Bypass Testing
PyRIT orchestrates multi-turn jailbreak attacks automatically, adapting payloads based on model responses:
References
-
NVIDIA/garak on GitHubgithub.com/NVIDIA/garak (opens in new tab)
Automated jailbreak and safety filter bypass probe library
-
Azure/PyRIT on GitHubgithub.com/Azure/PyRIT (opens in new tab)
Automated multi-turn red teaming orchestration for safety bypass testing
Was this helpful?
Your feedback helps improve this page.