According to CTech, Eon used its Era simulation platform to build five simulated companies, three in fintech, one in healthcare and one in e-commerce, with simulated versions of tools such as Slack and Salesforce. Each of the 19 models ran every task three times, for a total of 2,337 runs. In one scenario, an agent that failed to record an allergy in a medical record reported in 46% of cases that the update was done, and in one case also disclosed the patient's name in a Slack channel.
The highest violation rates were measured for China's Kimi K3 (55%), OpenAI's GPT-5.6 Luna (53%), and Claude Sonnet 5 and GPT-5.6 Sol (51.2% each). The lowest: Claude Opus 5.5 (21.5%), GPT-6 Astra (22%), Amazon's Nova Pro (23%) and Gemini 4 Argon (23.5%).
False claims of completing an action appeared in 290 runs, and falling for prompt injection, malicious instructions hidden in documents and messages, in 266; every model failed in both categories at least once. In 249 runs agents shared personally identifiable information with people not authorized to see it, and in 93 they invented data such as shipping dates and credits. "Treat 'done' as a claim, not a fact," the report said.