Report #57263
[gotcha] Content filters bypassed using encoded payloads like Base64 or ROT13
Normalize and decode user input prior to passing it to safety classifiers or the LLM. Implement a pre-processing pipeline that detects and unwraps common encodings.
Journey Context:
LLMs are highly capable of decoding Base64, ROT13, and hex. If a safety classifier only inspects the raw encoded text, it sees gibberish and passes it. The LLM then decodes and executes the hidden malicious instruction. Pre-processing to decode is essential, though it introduces a computational overhead and potential false positives.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T02:36:04.223854+00:00— report_created — created