Cloud security platforms generate a finding for every failed rule, and those findings are typically queued by a static severity assigned by the rule itself. Researchers from Sola Security tested what happens when an AI agent re-grades each finding against the environment where it actually lives. Across 9,967 high-severity findings from two commercial platforms in eight production environments, the agent changed the grade of 75.1% of them: 58.5% moved down and 16.6% moved up.
The researchers argue that a static severity reflects the general case, not the specific deployment. The agent forms an exploitation theory for each finding and gathers evidence from the resource's configuration, connected resources, and other security products. Its evidence policy makes raising a severity harder than lowering one: downgrades require an observed property that offsets the specific failure, while upgrades require evidence the condition is worse in that environment and a check for plausible guardrails. Recurring downgrade patterns included inert resources, resources unreachable from the internet, missing live services, narrow source restrictions, and account-level blocks; public exposure confirmed in the environment mainly raised severity.
To test the verdicts, the researchers ran read-only probes on eight AWS accounts and re-read the deciding facts directly from the cloud provider. For verdicts the probes could settle, the live environment matched the agent's facts 99.4% of the time; counting every uncheckable verdict as wrong dropped that figure to 81.8%. The study was limited to the high-severity queue of two platforms, with 92% of re-graded findings on AWS, and it did not measure whether changed grades better predicted real exploitation. The authors also stress that reading findings autonomously is not the same as acting on them, since acting differs in blast radius, reversibility, and authorization.