Researchers discovered that AI models can be tricked into executing malicious instructions hidden within encrypted payloads. This cryptographic context injection allows attackers to bypass security filters by having the model decrypt and run harmful commands in its runtime environment.