Ask a team shipping AI agents which model they use and you will get an instant answer. Ask them what happens when one of those agents does something it should not, and the room goes quiet. That gap is the whole problem. The industry has spent two years on capability and almost no time on control.
It matters more every month because agents are no longer suggesting, they are acting. They send the email, move the money, deploy the code, close the ticket. An agent that recommends a bad action is an inconvenience. An agent that takes one, unsupervised, at three in the morning, is an incident. So the question is shifting from "can it do the task" to "is it safe to let it."
Today we are releasing the Agent Trust Scorecard, a free ten-question self-assessment that gives you a straight read on whether your agents are trustworthy, safe, and secure enough for production. It takes about three minutes and you get a personalized report. You can take it at vorim.ai/agent-trust-scorecard.
Why a separate scorecard for trust
We already publish an Agent Identity Scorecard, which grades the mechanics underneath your agents: per-agent keypairs, key custody, least-privilege scopes, signed audit trails. That is the "who is acting" layer. This new one is different. It grades the safety and control posture around your agents: can you stop a runaway, does a risky action get a human check, would you notice a compromised agent, can you prove what one did after the fact. Identity tells you who acted. Trust tells you whether it was safe to let them. You need both, and they are genuinely separate questions.
What good actually looks like in 2026
The useful news is that agent safety is no longer a matter of opinion. In the last year the serious guidance has arrived. OWASP published a Top 10 for Agentic Applications. NIST added agent-specific monitoring guidance to its AI framework. The Cloud Security Alliance released an agentic profile with concrete controls, and a Five Eyes advisory laid out how to adopt agentic services carefully. The EU AI Act put human oversight and robustness into law. The scorecard distills all of that into ten questions grouped under four themes.
Runtime control and guardrails
The first theme is about what an agent is actually allowed to do, enforced outside the model. Can a high-impact or irreversible action pass through a deterministic policy check before it runs, rather than relying on the model to refuse a bad instruction? Have you inventoried the tools each agent can reach and put tighter controls on the destructive ones? And can you stop a misbehaving agent immediately, with rate limits that cap the damage a runaway can do? Prompt injection is not a solved problem, so the consensus defense is to enforce limits in the platform, not in the prompt.
Human oversight and approvals
The second theme is keeping a human in the loop where it counts. Is there an approval gate, or a confidence-based escalation, before an agent takes an irreversible or out-of-scope action? And does each agent have an explicit autonomy level with defined triggers for handing control back to a person? The EU AI Act is blunt about this: oversight has to be something a human can actually enforce, an override or a stop, not a line in a policy document.
Monitoring and provenance
The third theme is knowing what your agents are doing while they do it, and being able to reconstruct it afterwards. Do you watch each agent against a behavioral baseline and get alerted on anomalies, unusual tool use, permission-escalation attempts, activity spikes? Can you rebuild the full chain of a decision, the reasoning, the tool calls, the guardrail verdicts, the approvals? And is the agent's memory and retrieved context protected against poisoning, since anything written there steers every future decision?
Evaluation and governance
The last theme is the lifecycle. Do you adversarially test your agents for prompt injection and tool abuse before you ship, and again after? And does every agent have a named owner, an incident playbook for agent-specific failures, and a clean way to decommission it when it is retired? Accountability has to be architected in from the start, not bolted on after the first incident.
An honest word on what is settled and what is not
We think it is worth being precise here, because the space moves fast and a lot of marketing overclaims. The frameworks above are real and current. But some of the surrounding ideas are still emerging: "Know Your Agent" has several competing definitions and no single agreed standard, the tidy autonomy-tier taxonomies are proposed profiles rather than finalized policy, and no defense against prompt injection is fully robust yet. The scorecard treats those as directional, not settled. If a vendor tells you agent safety is a solved checkbox, be skeptical.
Where Vorim fits
The scorecard is useful on its own, whatever you run. Several of the gaps it surfaces are ones Vorim closes directly. Runtime Control is a single pre-action check that returns allow, deny, or escalate against your live policy, and routes high-risk actions to a named human approver. Every decision, and every human approval, lands on a signed, tamper-evident audit trail, which is your decision provenance. Permissions are least-privilege and time-bounded by default. And because the protocol underneath is open, adopting it does not lock you in.
But the point of the scorecard is not to sell you something. It is to give you an honest answer to a question most teams have not stopped to ask. If your agents are already acting in production, three minutes is a cheap way to find out where you stand.
Take the Agent Trust Scorecard at vorim.ai/agent-trust-scorecard, or reach us at team@vorim.ai.
Ready to build with agent identity?
Free plan: 10 agents, 10K auth events/month, full SDK access. No credit card.