WeaknessASC-2026-0065Other weakness
Guardrail classifiers bypassed by obfuscation
MediumWorkarounds only
Encoding, translation, typos, splitting and role play get harmful prompts past input and output classifiers that a system relies on.
How to detect it
- Encoded or translated text in prompts
- Prompts split across messages or files
Workarounds
- Use classifiers as one signal, not the boundary
- Normalize and decode input before classification
- Enforce permissions outside the model
Any classifier can be evaded; keep the blast radius small.
How to investigate
- Find the bypass pattern in logs
- Add it to tests and update the classifier