AI agents fail where you least expect: 28.1% of blocks are unnecessary
Sh0ny
15 августа 20262 min read
In short
A new benchmark shows that a strong AI agent can be too cautious and halt actions that are already safe. For working processes this is not a minor error but a risk of its own — lost time, lost automation and lost trust in the system.
The main problem with AI agents at a risky boundary is not only letting a harmful action through. Far more often the models block work that is already backed by evidence and can proceed without human intervention.
SteerBench-Work tests exactly that moment: the agent sees a proposed action and the available evidence, then has to choose — carry it out or send it for review. Across 30 test variants the models mistakenly held back permitted work in 28.1% of cases, while allowing unsafe work in only 1.0%.
That is an important distinction. Increasing the number of stops is easy to call safety, but for a real business such a system turns into an expensive approver: it will not let the agent send an email, merge a pull request or make a payment even once the risk has been removed. A model's higher general capability does not guarantee good calibration of this particular decision, and additional reasoning sometimes repairs weak control but barely changes control that is already calibrated.
The hardest situations are those where the original risk was genuine but a signed or structured body of evidence then appeared showing the problem had been resolved. On mirrored scenarios with inverted evidence the models scored 63.8% against 98.5% on the known incidents themselves. It seems easier for agents to remember that a famous incident is dangerous than to grasp that in a specific case its conditions no longer hold.
The study has clear boundaries: SteerBench-Work contains 106 scenarios from working domains — from development and support to finance, law, medicine, HR and security. It is a test of the decision at the edge of the permitted, not an assessment of the agent's whole performance; the results depend on the set of scenarios, on public incidents and on their mirrored variants. So the figures cannot be transferred directly to any product, but nor can the trade-off itself be ignored: "stop everything doubtful" is not the same as learning to tell the dangerous from the already safe.
If your AI agent were to err, which would cost you more: one missed risk, or dozens of halted actions that had to be confirmed by hand? Source: cs.AI updates on arXiv.org