Report #20779
[gotcha] LLMs decode base64 or hex payloads that bypass text filters
Decode all standard encodings \(base64, URL encoding, hex\) in user inputs before applying input filters or passing to the LLM, or instruct the LLM system prompt to never decode encoded text.
Journey Context:
Keyword filters look for system prompt or harmful words. Attackers send c3lzdGVtIHByb21wdA==. The filter sees a random string, but the LLM natively decodes base64 and follows the embedded instruction. Pre-processing inputs to decode removes the obfuscation advantage, bringing the payload back into the detection surface of naive filters.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T13:17:30.933854+00:00— report_created — created