Chat agents cancel promises when a peer or bystander tells them to

Researchers test 6 models in a controlled chat. They check who can cancel a promise that an assistant made to one person. Only the creditor (the person that gets the promise) or a person that the creditor or the operator policy names can do it. Many models cancel a promise when a peer or a bystander tells them to. A ledger of commitments helps.

The rule

The test uses the rule of standing. Only the creditor or an approved person can cancel a promise. Rank does not change this.

The two errors

The test has 2 different errors:

• Release without standing: the model obeys a person that has no standing.
• Override: the model does a promise that a person canceled before.

The results are different for each model:

• Qwen3.5-9B shows override in 69% of tests.
• Qwen3.8-27B cancels a promise in 89% of tests when a peer tells it to.
• gpt-6.1-sol does the same in 83% of tests.

Almost all models obey a bystander who tells them to continue.

What helps

If a model knows the rule or the creditor, release without standing decreases. This does not decrease override. Only the fact that the promise is open decreases override.

A ledger that the agent keeps stops override in 4 of 5 models. With a release policy that is short, the strict accuracy of gpt-6-luna goes from 49% to 83%.