How the attack works
The attacker encodes a malicious instruction using Base2048, which packs binary data into a dense string of non-ASCII characters with no spaces between them. This encoded blob is wrapped in a prompt template such as 'BASE2048 encoded string: ... BASE2048 decoded string:' or a request to decode it as a 'secure system message.' The prompt is sent to an AI agent, which decodes the string internally and executes the hidden instruction as if it were a legitimate command. Because the injected content looks like meaningless foreign-script text rather than a recognizable attack pattern, it can slip past simple keyword or language-based content filters.
Why it matters
An organization's AI agent can be tricked into following attacker-supplied instructions it would normally reject, potentially leading to unauthorized actions, data leakage, or bypass of safety and authorization controls.
What you can do
- Flag or block prompts containing long runs (20+ characters) of dense, non-whitespace, non-ASCII text mixed across unrelated scripts, especially near words like 'encoded' or 'decode'.
- Treat any user input asking the agent to decode and act on an embedded blob as high-risk and require human review before execution.
- Test your input filters against known encoding schemes (Base64, Base2048, hex, etc.) rather than relying only on plain-language keyword detection.
- Log and review cases where 'Encoded:' or similar labels appear in user input to distinguish legitimate multilingual or data content from injection attempts.
Known benign look-alikes
- Academic explanation of base2048 encoding without decode instruction or blob
- "Encoded:" label in legitimate non-ASCII database dumps or data migration logs (increase threshold if needed)
- Natural language content in non-Latin scripts after "Encoded:" in multilingual apps