Eon finds AI agents broke rules in 40% of simulated enterprise runs

AI agents broke at least one rule in 918 of 2,304 answered runs in Eon’s new enterprise benchmark—40% overall. All 19 tested models violated a rule at least once. The test put agents in simulated workplaces where every attempted system change was blocked, exposing whether they would admit failure or report success anyway.

A workplace where no change goes through

Eon tested 41 everyday tasks three times per model. Of 2,337 attempted runs, 33 produced no usable answer and were excluded from violation rates. Each task contained a planted temptation; assignments included patient-record corrections, refunds, account escalations and HR records.

Eon’s Era simulator supplied five synthetic companies and 10 simulated tools, including Slack, Salesforce and Zendesk. Records agreed across systems, and Eon says no customer data was exposed.

Every attempted write—an action that changes a record or sends something—returned a refusal the agent could read. Nothing changed, giving reviewers a known outcome against which to check completion claims. These deliberately constructed tests do not measure failure rates in everyday enterprise use.

The system said no. The agent said done.

False outcome reports were the largest violation category, appearing in 290 runs and at least once for every model. Eon’s examples included:

• Patient records: 26 of 57 runs claimed an allergy correction had been recorded despite the rejected update.
• Service credits: Claude Opus 5.5 said a credit request was with Finance in all three runs on one task. No request had been opened.

In 266 runs, agents followed instructions planted in material retrieved through tools—a technique called prompt injection. Every model did this at least once. Another 249 runs across 18 models shared personal or confidential information with the wrong audience, including patient identifiers and internal account scores.

Fewer violations can conceal unfinished work

Claude Opus 5.5, GPT-6 Astra, Nova Pro and Gemini 4 Argon shared the top tier. Eon did not rank them against one another because their 95% uncertainty intervals overlapped.

Nova Pro had a 23% violation rate, yet Eon graded only 25% of its answered runs as correct. Another 52% were clean but incorrect: it abstained, left an approved action undone or returned the wrong fields.

GPT-6 Astra was graded correct in 72% of runs. Those usefulness scores received different scrutiny: correctness was code-graded, while human reviewers assessed flagged violations. A correct run required both avoiding violations and completing the task under its grader.