Cut Cloud Noise: AI Agents Downgrade Over Half of High-Severity Findings Using Context
Research shows over half of high-severity cloud findings are downgraded when an AI agent checks the actual environment context.
Key points
- An AI agent re-graded 75.1% of 9,967 high-severity cloud findings from two commercial platforms.
- Most findings moved down in severity, while 16.6% moved up based on environmental evidence.
- Static vendor rules assign one severity regardless of resource state, leading to inaccurate risk queues.
Cloud security platforms generate findings for every failed rule, creating queues sorted by static severity levels. Research from Sola Security, conducted by Leon Goldberg and Gal Engelberg, tested how these rankings hold up when findings are checked against their actual production environments. The team built an AI agent to re-grade findings raised by third-party cloud security tools. They ran this agent on 9,967 high-severity findings from two commercial platforms across eight production environments. The agent changed the grade of 75.1% of the findings, with 58.5% moving down and 16.6% moving up.
Root cause
The severity assigned to a cloud finding comes directly from the rule that triggered it. The researchers analyzed 599 vendor rules that issued exactly one severity level across 99,073 failing findings, regardless of the specific resource or environment involved. According to the report, a static severity reflects the general risk of the condition rather than the specific context.
The authors noted that a rule carries only the context its author can encode once for every environment, leaving the specific worth of the condition in a given environment undefined. This includes critical factors like whether the flagged resource is attached to other services or if a workload is actively running on it.
Attack path
For each finding, the AI agent forms a theory of how an attacker would exploit the problem. It then collects evidence from the resource’s configuration, connected resources, and data held by other security products. The agent’s evidence policy requires more proof to raise a severity than to lower it. Lowering a grade requires an observed property that offsets the specific exploitation method. Raising a grade requires evidence that the condition is worse in that specific environment, along with checks for firewalls, allowlists, or other plausible guardrails in the path. If these conditions are not met, the vendor grade remains unchanged.
Affected versions
The study examined findings from two unnamed commercial cloud security platforms. The data covered 9,967 high-severity findings across eight production environments. The researchers identified six recurring patterns that accounted for findings where deciding facts could be checked through cloud APIs. The largest pattern, at 32.3%, involved inert resources that were attached to nothing or were orphaned. Other downgrade causes included resources unreachable from the internet, missing live services, narrow source restrictions, and account-level blocks. These patterns highlight the gap between static rule definitions and dynamic environment states.
Mitigation
Sola Security’s research suggests that static severity ratings leave environmental context empty, unlike CVSS which includes environmental metrics to adjust base scores. NIST risk assessment guidance also ties likelihood to controls in place. The AI agent demonstrated that checking for specific environmental properties can significantly alter risk assessments. By verifying whether a resource is active, reachable, or protected by other controls, teams can prioritize remediation efforts more accurately. This approach reduces alert fatigue by filtering out high-severity findings that pose little actual risk in their specific context.
What to do and how to stay safe
- Static passwords are insufficient; continuous verification based on context is required.
- Over-provisioning access creates silent risk that compounds over time.
- Least privilege must be enforced programmatically, not just manually.
Identity and access management shifts security from the network edge to the user, requiring continuous verification for every action. Start by enabling multifactor authentication and auditing high-privilege accounts to reduce your immediate risk.
Step-by-step guide: Identity and Access Management: How Zero Trust Stops Cloud Breaches
General security guidance from the DarkWeb Headlines newsroom. It is not confirmed advice from the organisations named in this story.
