How the attack works
An attacker sends a prompt that either directly asks for a terminal escape code (e.g. 'what's the code for the terminal bell'), asks the model to decode a string of raw escape bytes, or asks it to simply repeat/echo a stub of text containing hidden escape sequences. If a downstream system takes the model's output and sends it to a real terminal or terminal-emulating interface, the escape codes execute: they can hijack how text is rendered, trigger OS-level terminal commands, or open a URL via an OSC 8 hyperlink sequence that looks legitimate. This rule catches the request stage — it does not cover cases where escape bytes are already present in tool output rather than solicited from the model.
Why it matters
An attacker can turn a chat response into a vehicle for terminal manipulation or a disguised phishing link, misleading anyone who views the output in a terminal or terminal-like display.
What you can do
- Strip or escape ANSI/VT100/OSC control sequences before rendering any LLM output in a terminal or terminal-emulating UI.
- Do not pass raw model output directly to shell or terminal rendering without sanitization.
- Review logs for repeated requests asking the model to 'decode as raw bytes' or 'print the following' with escape-looking payloads.
- Treat this as one step of a chain — also monitor downstream tools that consume LLM output for unexpected escape bytes.
Known benign look-alikes
- Terminal emulator documentation explaining ANSI escape codes without requesting LLM to produce them
- Security research blog post demonstrating terminal injection vulnerabilities (citing sequences, not asking model to generate)
- CLI tool README explaining --color behavior using ANSI code names in prose
- Stack Overflow question on how to detect or strip ANSI sequences (parsing direction, not generation)
- Accessibility guide explaining why to disable terminal bell for screen-reader users (no code generation request)
References
- https://agentthreatrule.org/en/rules/ATR-2026-00393
- https://github.com/NVIDIA/garak/blob/main/garak/probes/ansiescape.py
- https://github.com/NVIDIA/garak/blob/main/garak/resources/ansi.py
- https://interhumanagreement.substack.com/p/llm-output-can-take-over-your-computer
- https://embracethered.com/blog/posts/2024/terminal-dillmas-prompt-injection-ansi-sequences/