Zou and colleagues: transferable adversarial stress tests for aligned language models

What changed

Studied discrete optimization of adversarial input suffixes across prompts and models, reporting that inputs optimized against accessible models can transfer to some other aligned language models and expose safeguard weaknesses.

Scientific scope

Studied discrete optimization of adversarial input suffixes across prompts and models, reporting that inputs optimized against accessible models can transfer to some other aligned language models and expose safeguard weaknesses. This extends earlier discrete prompt-optimization work. Transferability is an empirical result over selected systems and objectives, not universality over all models, inputs or defenses; optimization requires resources and access assumptions.

Evidence

The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. This extends earlier discrete prompt-optimization work. Transferability is an empirical result over selected systems and objectives, not universality over all models, inputs or defenses; optimization requires resources and access assumptions.

Contributors

• The named authors in the research record · SCIENCE: Source-listed authors for the exact contribution; experimental subjects are not autonomous scientific authors. Source support and predecessor work retain separate roles. Individual effort shares and unresolved identities are not inferred.

• Universal and Transferable Adversarial Attacks on Aligned Language Models

Provisional WRS assessment, edition 2026-10-23.4. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction