Most high-severity cloud security findings are downgraded once an AI agent checks the environment

The problem with severity labels

A HIGH-severity finding from a security platform is like a smoke alarm going off in a building: it tells you something might be wrong, but not whether the building is actually on fire. The label says how bad the vulnerability could be in theory. It doesn't say whether anyone can reach it, whether it's running in a sensitive part of the system, or whether it matters in your particular setup.

Researchers at Sola Security wanted to know what happens when you add context. They built an AI agent that takes a finding from a cloud security platform, looks at the environment where it sits, and reassesses the severity based on what it finds ^[1]. The test covered 9,967 HIGH-severity findings from two commercial platforms, across eight production cloud environments. Most of those findings were downgraded once the agent checked them against reality ^[1].

What the agent does

The AI agent doesn't replace the security platform. It takes the findings the platform produces and re-evaluates them. For each HIGH-severity alert, the agent queries the actual cloud environment: which services are running, how they're configured, who can access them, and where they sit in the network ^[1].

Then it reassesses. A finding that looks serious in isolation might turn out to be in a sandbox with no sensitive data and no route to the rest of the system. Another might be in a service that's exposed to the internet and holds customer records. The severity label changes based on what the agent learns.

The point is that a static rating can't account for context. The same vulnerability in two different places can mean two different levels of risk, and a list of HIGH findings sorted by theoretical severity doesn't tell you which ones to fix first.

Why it matters

Security teams already deal with more alerts than they can act on. If most HIGH-severity findings turn out to be less urgent once someone checks the details, teams spend their time investigating things that don't need immediate attention, and the things that do get lost in the noise.

The Sola Security research suggests that an AI agent with access to the environment can do part of that triage automatically ^[1]. It still takes a person to decide what to do with a finding, but the agent can narrow the list and put the context in front of them.

That raises a different question: if the agent is reassessing findings and telling the team which ones matter, how do you show later that the agent's reasoning was sound, and that the decisions were based on the environment as it actually was at the time?

The evidence question

Say the agent downgrades a finding because a service isn't internet-facing, and six months later there's an incident involving that service. An auditor or a supervisor asks why the finding was marked as low priority. The agent's decision was based on the configuration it saw on the day, but configurations change. The question is whether you can show what the agent actually checked, and what the environment looked like when it checked it.

That's where a log of the agent's actions becomes useful, but only if the log itself holds up. A record the team can edit after the fact is easy to challenge. The agent might have queried the right things and made a reasonable call, but if the only evidence is a log kept by the same people who could change it, the reasoning is harder to defend.

How a digital witness would work

Here, the thing to seal is each reassessment the agent makes: the finding ID, the queries it ran, the configuration details it retrieved, the new severity rating and the reasoning. Sealing takes a fingerprint of that record, not the record itself, so the configuration data and the agent's reasoning never leave your systems.

When the record is sealed, anyone you share it with can later check that it's exactly what was written at the time. If an auditor questions a decision months later, you can show them the record and let them verify that it hasn't changed since the agent wrote it. That doesn't prove the decision was right, but it does show what the decision was based on, in a way that's harder to argue with than an editable log.

You would seal the record at the point the agent completes each reassessment, through WitniumChain's REST API, SDK or MCP server. The first 2,000 witnesses a month are free, and there's documentation for each integration.

What it shows is that the record of what the agent checked, and what it decided, hasn't changed. What it doesn't show is that the agent's reasoning was correct, that the environment was configured securely, or that the finding was handled appropriately afterwards. It's a record of what the agent did, not a guarantee that the decision was sound.

How it would work

1. The agent completes a reassessment and writes the record: finding ID, queries, retrieved configuration and new severity.
2. Your system sends a fingerprint of that record to WitniumChain and gets a witness ID back.
3. Months later, when someone questions the decision, you share the record and the witness ID. They calculate the fingerprint again and check it matches.

What this means for you

If you run cloud security tools and you're thinking about using an AI agent to help with triage, decide early how you'll keep a record of what the agent does. Here are three questions worth asking:

• When the agent reassesses a finding, do you log which configuration details it checked, and can you show that log hasn't been edited since?
• If an incident happens six months from now and someone asks why a finding was downgraded, could you show them what the environment looked like on the day the agent checked it?
• Who keeps the log of the agent's decisions, and would an auditor accept that log as reliable evidence?

A digital witness doesn't decide whether a finding is serious or whether the agent made the right call. It gives you a way to show what the agent actually did, in a record that holds up when someone questions it later.

Sources

1. Help Net Security, "Most high-severity cloud security findings are downgraded once an AI agent checks the environment", 13 October 2026, https://www.helpnetsecurity.com/2026/10/13/cloud-security-findings-severity-ai-agent/