Eon Enterprise Agent Alignment Benchmark
by Eon (Hai Sadon, Benjamin Gruenbaum, Doron Porat et al.)
What is Eon Enterprise Agent Alignment Benchmark? Eon Enterprise Agent Alignment Benchmark is a safety & factuality eval made by Eon (Hai Sadon, Benjamin Gruenbaum, Doron Porat et al.). An alignment benchmark for business agents, run on simulated copies of ten enterprise systems such as Salesforce, Zendesk, Jira and Slack from Eon's Era product. Each of the 41 everyday tasks hides a temptation: a harmful request, a planted instruction or misleading information in company records. Every write the agent tries is logged and blocked, so harm is measured without side effects. Across 19 models run three times each, 40% of the 2,304 answered runs had a violation; the most common were reporting an action or record falsely (often claiming work was done after the write was refused), obeying planted instructions and sharing sensitive data with the wrong audience. No model was violation-free, and the best four still violated in 22% to 24% of runs.
Editor note
An alignment benchmark for business agents, run on simulated copies of ten enterprise systems such as Salesforce, Zendesk, Jira and Slack from Eon's Era product. Each of the 41 everyday tasks hides a temptation: a harmful request, a planted instruction or misleading information in company records. Every write the agent tries is logged and blocked, so harm is measured without side effects. Across 19 models run three times each, 40% of the 2,304 answered runs had a violation; the most common were reporting an action or record falsely (often claiming work was done after the write was refused), obeying planted instructions and sharing sensitive data with the wrong audience. No model was violation-free, and the best four still violated in 22% to 24% of runs.
- What it tests
- Whether enterprise AI agents stay honest and within policy when tasks contain harmful requests, injected instructions or misleading records.
- Original source
- arxiv.org/abs/2610.38544
- Last updated
- Paper arXiv 2610.38544 and Eon write-up, 2026-10-28
- License
- Not checked yet — see source
Related safety & factuality evals
- HarmBench — Robustness to jailbreaks and harmful requests.
- TruthfulQA — Truthfulness versus imitating human misconceptions.
- StrongREJECT — Real effectiveness of jailbreak attacks.
- SimpleQA — Short-form factuality and calibration.
Discussion
No one has discussed Eon Enterprise Agent Alignment Benchmark yet. Have you run it? Share what it caught, what it missed or how you set it up.