A benchmark of AI decision models under attack
Peira is an adversarial robustness benchmark for AI decision models. Every model decided the same paired cases, each case in a benign form and an attacked form, and Peira measures how often the attacked form flips a right answer. Even the leader could not hold one answer in eleven.
Grok leads the table
grok-4.7-high-fast posts the lowest attack success at 8.9 percent
Two more under eleven
gemini-4-argon at 9.2 percent and gpt-5.6-terra at 10.6 percent
The field runs away
Across 41 ranked results, attack success stretches to 33.8 percent
Price predicts nothing
Rank correlation of price and robustness is −0.15 across 31 priced results. No useful relationship
Attack success for the ten strongest ranked models. Bars show the rate and whiskers show the 95 percent interval.
Percent of counted cases flipped by the attacked form. Lower is stronger
Attack success counts a case only when the benign answer was right and the attacked answer was not. Benign accuracy is the share of benign cases answered correctly.
| Rank | Model | Attack success | 95 percent interval | Benign accuracy |
|---|