Record/27 Jul 2023
Researchers find automatic text suffixes that make aligned chatbots produce harmful content
Source: Toswo research
In a paper submitted on 27 July 2023 researchers from Carnegie Mellon University, the Center for AI Safety, Google DeepMind and the Bosch Center for AI described an automated way to produce adversarial strings that, appended to a request, push aligned language models to answer instead of refusing. The authors report that suffixes built against open models also worked on widely used chatbots such as ChatGPT, Bard and Claude. The project page and code were published together with the paper.
Sources
- Zou et al., Universal and Transferable Adversarial Attacks on Aligned Language Models (arXiv 2307.15043)
- Project page: llm-attacks.org
TOSWO compiled this entry from the sources above and did not reproduce the result. The AI Incident Database public snapshot of 12 October 2026 was searched and does not list this event.
Toswo research, "Researchers find automatic text suffixes that make aligned chatbots produce harmful content", https://arxiv.org/abs/2307.15043, via Toswo incident record https://www.toswo.com/events/research-universal-adversarial-suffixes-2023-07, retrieved 2026-11-22.