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
Limited correlation between robustness and model size, capability or inference-time compute.
More capable models were often more susceptible to tool poisoning.
Extended thinking raised ART attack success for Opus 4.6 (21.7% vs 14.8% at k=100); Anthropic says the effect did not replicate on its other evaluations.
Cuts both ways: Opus 5.5 has lower measured attack success than Opus 5 on the Gray Swan benchmark (1.0% against 4.8% at k=15), while early snapshots followed planted instructions in user-pasted text in 52% of attempts, a regression Anthropic traces to training data.