Research
Outsmarting AI Guardrails with Invisible Characters and Adversarial Prompts
Publication date unknown · Discovered mindgard.ai
Page published
Publication date unknown · First observed: 25 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Mindgard's findings show why defenders should test guardrails against Unicode obfuscation and adversarial perturbations rather than rely on prompt-injection detection as a standalone security boundary.
Mindgard reports an empirical evaluation of six LLM guardrails, finding weaknesses against character obfuscation and adversarial perturbations. The study reports complete detection evasion for some techniques, including emoji smuggling, across several guardrails, including Protect AI v2 and Azure Prompt Shield; these results concern detector bypass rather than demonstrated downstream compromise.