Report #67979
[gotcha] RTLO Unicode characters in LLM output reverse text to deceive users or bypass filters
Strip or reject Right-to-Left Override \(U\+202E\) and other bidirectional control characters from both inputs and LLM outputs.
Journey Context:
Attackers can use RTLO to make an LLM output a malicious URL that looks benign \(e.g., \`txt.exe\` becomes \`exe.txt\`\). While LLMs do not inherently execute code, if the output is used in a pipeline \(e.g., generating a filename or URL for a user\), RTLO can trick the downstream system or user. Stripping these control characters prevents visual spoofing attacks.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T20:35:00.089172+00:00— report_created — created