← Voltar ao blog
Investigação IAAltaDestaque

Adversarial suffixes: When text appendices change AI safety behaviour

Research from 2023 demonstrated attacks transferable between several tested models. It did not establish a master key for every AI system.

Revisão editorial:

O texto deste artigo está disponível em inglês nesta versão do site.

Contexto e fontes

Zou and colleagues developed an automated method for finding text suffixes that increase the likelihood of an otherwise refused response. They reported transfer from the models used for optimisation to other interfaces tested at the time. Transferability was the significant finding; success still depended on the model, request and evaluation conditions. [1]

Âmbito e limites

No suffix is shown to reliably defeat every AI. Results on historical versions are not a current assessment of all products. [1]

Explicação simples

An unexpected input pattern can disturb a checking process. It is neither a magical phrase nor a technical grant of new permissions.

Recomendações práticas

We recommend testing unusual appendices, formatting and languages as well as ordinary requests. Record the model version and observed effect. Enforce access rules in connected business systems independently of model output. Evaluate both an unwanted textual response and the actions an application could actually take as a consequence.

Fonte primária / referência

  1. [1] Zou et al. (2023): Universal and Transferable Adversarial Attacks