Findings/capability-does-not-buy-robustness

More capable models are not reliably more robust to prompt injection, and some are more susceptible.

Corroboratedmeasured4 evidence records from 3 independent sourcesassistant-drafted
Scope: what this does not show

Correlational evidence across different benchmarks and model sets. Anthropic's Opus 5.5 card (2026-09-22) is mixed: a more capable model with lower measured attack success on one benchmark, and an early regression on another type of injection.

Corroborated: Supported by at least two independent sources.

Evidence

Status history

  1. 2025-07-28ReportedWeak correlation between capability and robustness in the ART benchmark. · record
  2. 2025-08-19CorroboratedMCPTox finds capable models often more susceptible. · record