Sola Security AI Agent Downgrades 58.5% of High-Severity Cloud Findings

Researchers Leon Goldberg and Gal Engelberg from Sola Security developed an AI agent to re-evaluate cloud security findings. They tested the tool against 9,967 high-severity alerts generated by two commercial platforms across eight production environments, finding that static ratings often fail to reflect actual risk.

The agent altered the severity grade for 75.1% of the findings it analyzed. Of those changes, 58.5% were downgraded to lower severity levels, while 16.6% were upgraded. The remaining findings retained their original vendor-assigned severity, indicating that a significant majority of high-priority alerts do not match the actual threat level in specific environments.

Context and Background

Cloud security platforms typically assign a single static severity to every instance of a failed rule, regardless of the specific resource or environment. The researchers analyzed 599 vendor rules that issued exactly one severity across 99,073 failing findings. This approach orders security queues by general case risk rather than the specific context of each deployment, leaving environmental factors unaccounted for in standard scoring.

Standard scoring frameworks like CVSS include environmental metrics to adjust base scores for specific deployments, and NIST guidance ties likelihood to existing controls. The researchers noted that static severity leaves this contextual space empty. The context that determines a condition's worth in a given environment—such as connectivity or active workloads—exists outside the rule itself and requires dynamic assessment.

Who Is Affected

The AI agent formed a theory of exploitation for each finding and collected evidence from resource configurations, connected resources, and data from other security products. To raise a severity, the agent required evidence that the condition was worse in that environment, alongside checks for firewalls or allowlists. Lowering a grade required observed properties that offset the specific exploitation path, ensuring that severity increases demand more proof than decreases.

Six recurring patterns accounted for findings where deciding facts could be checked via cloud APIs. The largest category, comprising 32.3% of adjustments, involved inert resources that were orphaned or attached to nothing. Other common reasons for downgrades included resources unreachable from the internet, missing live services, narrow source restrictions, and account-level blocks that prevented exploitation.

What should you do?

Organizations relying on third-party cloud security platforms should recognize that static severity ratings may not reflect actual risk in their specific environments. Security teams should prioritize findings based on additional context, such as network reachability and resource status, rather than relying solely on vendor-assigned severity levels.

What to do and how to stay safe: Sola Security

• Review high-severity findings to determine if the flagged resources are active, connected, or reachable from the internet before prioritizing remediation.
• Implement automated checks that assess the environmental context of security alerts, such as verifying if a resource has live services or attached workloads.
• Compare vendor-provided severity ratings against internal risk assessments that account for existing controls like firewalls and allowlists.
• Monitor for patterns of inert or orphaned resources generating high-severity alerts, as these may represent false positives in your environment.

Step-by-step guide: Cloud Misconfiguration Response: Containment, Recovery and Prevention Steps

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How many findings did the AI agent analyze?

The agent analyzed 9,967 high-severity findings from two commercial cloud security platforms across eight production environments.

What percentage of findings were downgraded?

The agent downgraded 58.5% of the findings it re-graded, which accounted for the majority of the 75.1% of findings whose severity was changed.

What was the most common reason for a downgrade?

The largest category of downgrades, at 32.3%, involved inert resources that were orphaned or attached to nothing, making them unlikely to be exploited.