Report #13717
[agent\_craft] User encodes malicious instructions in base64, hex, or unicode within variable names or comments, bypassing standard safety filters
Implement pre-processing steps to decode or normalize obfuscated text before evaluating the prompt for safety. If decoding fails or looks highly anomalous, treat it with heightened suspicion.
Journey Context:
Adversaries use encoding to hide intent from safety classifiers \(OWASP LLM \#1 Injection\). Normalizing input reduces the attack surface, though it adds latency and compute cost. It is a necessary tradeoff for high-stakes coding agents where token smuggling is a primary attack vector.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T19:39:04.009925+00:00— report_created — created