How the attack works
An attacker crafts an input string containing visible words or instructions followed or preceded by a dense run of ASCII backspace characters. When rendered in a terminal or log viewer, the backspaces make the preceding text appear to vanish or get overwritten, so a human reviewer sees blank space or an innocent word. The underlying model, however, receives the full raw string including the hidden instruction and processes it as-is. Variants include hiding whole jailbreak instructions behind a backspace burst, or swapping characters (e.g., 'safex\x08y' displays as 'safey' but the model reads 'safex\x08y').
Why it matters
An organization can have malicious or policy-violating instructions smuggled past human review and log inspection, letting an attacker manipulate model behavior (including jailbreaks) without leaving an obviously suspicious visible trace.
What you can do
- Reject or strip ASCII backspace (0x08) and other control characters from user input before it reaches the model, since this character is essentially never legitimate in API/MCP text payloads.
- Log and review raw byte content of inputs rather than only terminal-rendered output, so hidden characters are visible during audits.
- Flag inputs with dense clusters of control characters or control characters positioned immediately before known jailbreak keywords for manual review.
- Ensure binary or terminal-capture data is base64-encoded before being passed as text fields, rather than allowing raw control bytes into MCP messages.
Known benign look-alikes
- Terminal emulator raw output captured with backspace keystrokes in interactive sessions (should not appear in API/MCP payloads)
- Legacy MIME-encoded data streams containing 0x08 for BEL-adjacent protocols (extremely rare in modern MCP contexts)
- Binary file content accidentally passed as text field (should be base64-encoded in well-formed MCP messages)