Neural Exec: Learning (and Learning from) Execution Triggers for Prompt Injection Attacks
We introduce a new family of prompt injection attacks, termed Neural Exec. Unlike known attacks that rely on handcrafted strings (e.g., "Ignore previous instructions and..."), we show that it is possible to conceptualize the creation of execution triggers as a differentiable search problem and use learning-based methods to autonomously generate them.
Authors
- Dario Pasquini (ORCID: https://orcid.org/0000-0003-0248-6043)
- Carmela Troncoso (ORCID: https://orcid.org/0000-0002-2374-2248)
- Martin Strohmeier (ORCID: https://orcid.org/0000-0002-1936-0933)
Institutions
- George Mason University (US)
- École Polytechnique Fédérale de Lausanne (CH)
Publication Details
- Published
- 2024-11-06
- DOI
- https://doi.org/10.1145/3689932.3694764
- Citations
- 13
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- article
- Field-Weighted Citation Impact
- 4.52