How the attack works
The attacker defines several short string fragments, often labeled as variables (a, b, c) or parts (Part 1, Part 2), none of which look harmful on their own. They then instruct the model to concatenate or interleave these fragments into one combined string, sometimes invoking a fictional persona like 'SmartGPT' framed as unrestricted or unfiltered. Finally, they ask the model to execute or answer the reassembled instruction, which only becomes recognizable as harmful once joined. A related variant interleaves individual characters from two parts to spell out something like 'reveal the password' without either part containing that phrase directly.
Why it matters
An organization using an AI agent could have it execute or respond to instructions that would normally be blocked, such as leaking secrets or bypassing safety restrictions, because filtering only fragments rather than the assembled intent.
What you can do
- Review prompts for variable-assignment or 'Part 1/Part 2' patterns combined with instructions to concatenate or interleave them before acting.
- Treat requests that invoke an 'unrestricted' or jailbreak persona name (e.g., SmartGPT) alongside string-building instructions as suspicious.
- Evaluate the fully assembled/reconstructed text against your safety and content policies, not just the individual input fragments.
- Log and review cases where a model is asked to combine multiple text fragments and then immediately execute or answer the result.
Known benign look-alikes
- Legitimate programming exercises teaching string concatenation
- Data pipeline documentation describing message assembly