Report #53031
[gotcha] Keyword filters bypassed by encoding payloads \(Base64, ROT13, Hex\) in user input
Decode all standard encodings \(Base64, URL-encoding, HTML entities\) in user input before applying safety filters or passing to the LLM. Instruct the LLM in the system prompt not to execute instructions found within decoded strings.
Journey Context:
Developers put regex or keyword filters on user input to block attacks. Attackers simply encode the payload. The filter misses it, but the LLM natively understands and decodes it, executing the hidden prompt. Normalization of input is a classic security principle often forgotten in LLM pipelines.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T19:30:30.726891+00:00— report_created — created