---
title: "Resolving the Enterprise: Unified Third-Party Findings in the Enterprise Security Brain - Sola Security"
canonical: "https://sola.security/insights/resolving-the-enterprise-unified-third-party-findings-in-the-enterprise-security-brain/"
---

# Resolving the Enterprise: Unified Third-Party Findings in the Enterprise Security Brain

[Keren Gaiger](https://sola.security/authors/keren-gaiger/)

|
Last updated: Nov 10, 2026

Enterprises rely on security tools for their domain expertise, but the context that decides how much a finding matters is spread across many systems: who owns the resource, what data it holds, and what it connects to. The Sola Security Brain closes that gap.

Enterprises choose security tools for their expertise in the areas that matter to their business. Each tool detects problems in its domain and judges them with the context it holds. The context that decides how much a finding matters to a particular organization is spread across many systems: what the resource does for the business, who owns it, what data it holds, what else in the environment it connects to, and what other tools have found on it.

The Sola Security Brain closes that gap. Enterprises bring the findings of the security tools they already use into the brain. Sola attaches each finding to the resource it concerns and reads it alongside everything else the brain knows about that resource: its business and organizational context, and the environment it lives in. High and critical findings on risky resources are investigated in that context, and their default severity is refined accordingly; we evaluated this severity refinement capability in a production field study [1]. When an enterprise works with several providers, the same brain gives it one place to see all of them: findings are normalized, so that findings describing the same underlying issue are recognized and each issue is counted once.

The approach is general. Adding a new security tool means describing how its findings map to the resources the brain already tracks; everything after that is shared.

## One brain for every provider

The Sola Security Brain is a graph of an organization’s security-relevant state [2], [3]. Alongside the resources and identities it tracks, it holds the security findings raised on them.

Enterprises also work with third-party security tools, chosen for their expertise in the areas that concern their business. That expertise is valuable, and Sola builds on it rather than replacing it. Each tool assigns severity using the context available to it. The brain adds a broader view: it combines the resource’s role in the business, the organization around it, the data it holds, the attack paths and blast radius that run through it, and the findings of every other connected tool.

Integrating a tool’s findings into the brain puts that expertise next to everything else Sola knows, so each finding can be read in the context of the specific environment its resource lives in. The main goal of the integration is to make the security telemetry of every connected tool accessible to agentic AI in a sound form: each finding attached to the resource it concerns, recognized when another tool reports the same issue, and graded in the context of its environment. An agent that reasons over the brain then works from resolved facts rather than from raw, overlapping vendor records. Three goals serve it.

1. **Contextualize tool findings.** Bring the findings and expertise of the security tools an enterprise works with into the brain, and refine their severity with business, organizational, and environmental context.
2. **See everything in one place.** When an enterprise works with several providers, show all of their findings together, normalized so that the same issue reported by several tools is recognized and counted once.
3. **Integrate any tool.** Make adding a new security tool a matter of describing its findings instead of building a new pipeline.

The scope is misconfigurations, the largest and least structured class of third-party findings.

## What stands between a finding and its context

Three challenges separate a raw tool finding from a contextualized fact in the brain.

**Context is spread across systems.** A tool’s default severity reflects how severe a misconfiguration usually is, together with whatever context that tool holds. Whether a public bucket serves a marketing site or holds customer exports, whether the resource belongs to a team handling regulated data, and whether it sits on an attack path to production are facts held in different systems. Two identical findings can matter very differently to the same organization, and telling them apart requires bringing those facts together.

**Findings have to land on the right resource.** Context lives on the resources the brain tracks, so a finding can be contextualized only once it is attached to the resource it concerns. Each tool reports findings against resources in its own data model, and working out how those records join to the brain’s resources takes careful analysis, tool by tool and resource type by resource type. The aim is to match as many findings as possible while never attaching a finding to an entity other than the one it lies on. A finding attached to the wrong resource plants a false fact in every question that touches that resource.

**Several providers report the same issue.** When an enterprise works with more than one tool, misconfigurations overlap. They form an open set without a shared identifier: each provider ships its own checks, and many describe the same underlying issue in different words or at a slightly different level of generality. A security group open to SSH from anywhere can be flagged by every connected provider, and the brain may already model the same issue as one of its own security controls. Counted naively, one exposure becomes several findings, and the resource’s risk grows with every provider the enterprise connects.

Identifier matching, duplicate recognition, and severity judgment are therefore done ahead of time and over the whole population. Every reader of the brain, from risk scoring to an investigating agent, then reads the answer instead of re-deriving it at question time.

## Three layers

Sola brings tool findings into the brain in three layers, each built on the one before it. Only the first is specific to a tool; normalization and contextual severity work the same way for every finding, whichever tool produced it.

**The first layer, resource matching,** attaches each provider finding to the resource the brain tracks. A failing provider check on a security group, for example, becomes an edge from that security group to the misconfiguration, so the layer adds misconfiguration nodes and resource edges to the graph. Matching is done by join rules, one per provider and resource type, each built in two stages.

1. **Research over live data.** For each provider and resource type, we analyzed how its findings reference resources in production data and how they join to the matching resources in the brain.
2. **Coverage and precision validation.** Each rule is checked against live data in both directions: findings that should have matched and did not, and matches that should not have happened. Guard conditions, such as matching the account, keep identical names in different accounts apart.

**The second layer, normalization,** recognizes findings, from any provider or from Sola’s own controls, that describe the same issue. “Unrestricted SSH access” and “Security group allows 0.0.0.0/0 on port 22” become one issue. The layer produces one cluster per unique issue, and risk aggregation reads the clusters rather than the individual findings.

**The third layer, contextual severity,** refines the default severity of high and critical findings on risky resources. A public bucket finding is raised or lowered depending on what the bucket holds and what it serves, and the refined severity carries the reasoning behind it.

## Normalizing findings: one issue, counted once

When an enterprise works with several providers, their findings sit side by side on the same resources, often alongside one of Sola’s own security controls. Seeing everything in one place is useful only if each issue appears once, whatever the number of tools that report it.

### What counts as the same issue

Two findings describe the same issue only if they describe the same security problem with the same scope and the same risk. The test is concrete: suppose one resource triggers both. They are duplicates only if dropping one of them loses no information. If dropping one would hide a distinct permission, scope, service, or condition, they stay separate (Table 1).

| Finding A | Finding B | Verdict |
| --- | --- | --- |
| Security group allows 0.0.0.0/0 on port 22 | Unrestricted SSH access | Same issue |
| Unencrypted EBS volume (Sola security control) | EBS volume without encryption at rest | Same issue |
| S3 bucket publicly readable | S3 bucket publicly writable | Different permission |
| Accessible by anonymous internet users | Accessible by any authenticated cloud user | Different access model |

Table 1. Same issue and near misses.

The rule is deliberately strict, because the two possible errors are asymmetric. Splitting two true duplicates counts a problem twice. Merging two distinct findings makes a real risk disappear from the count. When in doubt, findings stay apart.

A provider finding that lands in the same cluster as a Sola control corroborates an issue the brain already models. One that lands in a cluster of its own extends the brain’s coverage with that provider’s expertise.

### How findings are grouped

Semantic similarity alone cannot apply this rule, since “publicly readable” and “publicly writable” read almost the same. An LLM alone cannot compare thousands of findings with one another. Each therefore does one part of the job: semantic search finds the candidates, and the LLM decides which candidates describe the same issue (Figure 1).

**Initial clustering.** All known findings, Sola’s own controls and every provider check seen so far, are grouped by semantic similarity into loose candidate pools within each resource type. An LLM then splits each pool into clusters of findings that describe the same issue.

**New findings.** When a new finding arrives, semantic search fetches the closest existing clusters as candidates. The LLM is shown the new finding next to the findings in those clusters and decides whether it describes the same issue as one of them; if it does, the finding joins that cluster, and otherwise it opens a new one. A new provider check titled “SSH port open to the internet” on a security group, for example, is matched to the cluster holding “Unrestricted SSH access” and is counted as the same issue from then on.

| Step | Initial clustering
All known findings, grouped once | New findings
Each new finding, as it arrives |
| --- | --- | --- |
| **Input** | **All known findings**
Sola security controls and every provider check seen so far. | **New finding**
“SSH port open to the internet” |
| **Semantic search** | **Candidate pools**
Findings close in meaning, on the same resource type. | **Candidate clusters**
The closest clusters, such as “Unrestricted SSH access”. |
| **LLM decision** | **Refine into clusters**
Each cluster holds findings that describe the same issue. | **Same issue?**
Join the matching cluster, or open a new one. |

Figure 1. Grouping findings into issues.

## Contextualizing findings: severity in context

High and critical misconfigurations on risky resources are investigated in the context of the environment they live in: the resource’s role, its business and organizational context, the data on it, and the attack vectors and blast radius that run through it. The investigation produces a refined severity and the reasoning behind it. A public bucket serving marketing assets can be lowered; the same finding on a bucket holding customer exports, reachable from an internet-facing workload, can be raised. We evaluated this severity refinement capability on third-party cloud security findings in production environments [1].

Provider findings take part in the investigation in two ways.

- **As evidence.** When a finding is investigated, the findings that other providers have raised on the same resource are part of the evidence. Several providers, each with different expertise, pointing at related weaknesses on one resource is a signal in its own right.
- **As subjects.** A provider finding that has no counterpart among Sola’s own controls is investigated on its own, so a check Sola never wrote still receives a contextual severity.

Because duplicates are already recognized, each issue is weighed once.

## What this enables

A brain that integrates, normalizes, and contextualizes tool findings supports four capabilities.

- **Tool expertise, in business context.** Each tool’s findings keep the expertise behind them and are read in the business, organizational, and environmental context the brain holds, with severities refined accordingly.
- **Everything in one place.** Enterprises that work with several providers see all of their findings together, alongside the brain’s own findings, on the resources they concern.
- **Risk that counts issues, not reports.** Normalization shows which findings describe the same issue. Each unique issue contributes once to risk aggregation, so risk does not inflate as tools are added.
- **Any tool, one pattern.** A new security tool is integrated by describing how its findings map to the brain’s resources. Normalization and context come with it.

## Key takeaways

- **Tools bring expertise; the brain adds context.** Each tool knows its domain. The brain brings together the business, the organization, and the environment around each resource, and uses them to refine a default severity.
- **Attach before you reason.** Context lives on resources, so a finding is useful only once it is on the right one. Join rules are validated against live data in both directions.
- **One place, each issue once.** With several providers, findings that describe the same issue, from any tool or from Sola’s own controls, are recognized as one. Embeddings propose candidates; strict LLM judgment decides.
- **Any tool, one pattern.** Only the mapping to resources is tool-specific. Normalization and contextualization are shared, so every new integration benefits from them immediately.
- **Every tool adds context, not noise.** Each new integration enriches the brain without inflating the risk it reports.

> When the brain counts a misconfiguration, it counts it once, on the resource it belongs to, at the severity it deserves.

---

## References

[1] L. Goldberg, G. Engelberg. [Contextualization of Third-Party Cloud Security Findings](https://arxiv.org/abs/2610.08895). arXiv:2610.08895, 2026.

[2] L. Goldberg, G. Engelberg, E. Yavin, E. Elouz, A. Zadok, K. Koutsyi. [Coding Agents Aren’t Enough! Evaluating an Enterprise Security Brain for Agentic Cloud Investigations](https://arxiv.org/abs/2609.30345). arXiv:2609.30345, 2026.

[3] L. Goldberg. Context is the Architecture: Rethinking Security Platforms in the LLM Era. Sola Security, July 2026.

About the author

[Keren Gaiger](https://sola.security/authors/keren-gaiger/)

Senior Applied AI Researcher

[![Keren Gaiger on LinkedIn](https://sola.security/wp-content/themes/sola/images/ico-linkedin.svg)](https://www.linkedin.com/in/keren-gaiger-711558100/)
