Why is this finding reachable? Asking Safeguard through ChatGPT

Connected to the Safeguard MCP server, ChatGPT can list your findings and fetch CVE details, which gives you the context to question a reachability verdict. Safeguard's documentation shows the reachability class itself in the platform's Vulnerabilities view and in its own search chat, so confirm the verdict there before acting on it.

We could not confirm from the documentation a dedicated MCP tool that returns call paths, so this post is honest about what a ChatGPT session can and cannot answer. If you want to connect first, see our setup steps in the MCP how-to.

What does "reachable" mean in Safeguard?

Safeguard's docs describe reachability as whether your code actually calls the vulnerable function or code path. They describe three kinds of analysis: static, which builds a call graph from source, lockfile and bytecode; dynamic, which observes what runs; and configuration, which checks whether the code is wired in at all.

Results fall into four classes:

• Reachable: a real call path exists and is instantiated at runtime.
• Conditionally reachable: a path exists but needs a feature flag, route or setting that is not enabled today.
• Unreachable: no call path exists in your compiled code.
• Unknown: the analyzer could not decide, for example because of reflection or native code. The docs say it is marked conservatively as potentially reachable.

The docs also say reachability is weakest with reflection, eval or unsafe dispatch, with data-driven flaws, and with code paths that depend on configuration the analyzer cannot enumerate. In those cases they point to other signals such as KEV, EPSS and runtime observation. Reachability changes priority, not truth.

What can ChatGPT see through the MCP server?

A new Safeguard workspace starts with six read-only tools: list products, list projects, get a project with its components and vulnerabilities, list vulnerabilities (filterable by severity, ecosystem, CVSS score or exploit status), get one CVE with its CVSS score and remediation, and list findings in an SBOM or project. Other tools are off until a tenant administrator enables them.

That is enough to ask questions like these:

1. "List findings for the checkout project, highest severity first."
2. "Which of these CVEs have a known exploit?"
3. "Show details for this CVE and the recommended fix."
4. "Why would a flaw in this package matter to a service that only uses its date helpers?"

The last one is reasoning, not a lookup. ChatGPT can explain how a library flaw usually triggers, using the CVE description the tool returned, but it is not reading your call graph unless a tool supplies it. Treat its answer as a hypothesis.

Where do you confirm the reachability class?

In the Safeguard web app. The docs say the Vulnerabilities tab of a project has a Priority column and a Reachability column, with filters such as reachable-only. They also say Safeguard's search chat understands the prioritization surface, with example questions like "Show my P0 findings in production that are reachable." Use that chat when you need the verdict, and use ChatGPT through MCP when you want the wider conversation around it.

Ask the follow-up that matters: reachable via what? You want the entry point, the call path to the vulnerable function, and whether the entry point is exposed to untrusted input.

How do you challenge a verdict?

• If it says unreachable, check for dynamic loading, reflection or plugin systems. Those are the cases static analysis handles worst.
• If it says reachable, find the entry point and confirm the vulnerable function is actually called with attacker-influenced data.
• If it says unknown, do not dismiss it. Treat it as potentially reachable until someone checks.
• If the CVE is on CISA's KEV list, exposure matters more than reachability labels. Known exploitation is a strong reason to act.

What are the limits?

The reachability analysis quality varies by language, and the docs mark some ecosystems as preview. An assistant on top of a read-only tool set can only reason about what the tools return. And any chat answer can be wrong, so keep the evidence in the platform and the decision with a human.

What to do

1. Connect ChatGPT to the Safeguard MCP server and check the tool list.
2. List findings for one project and note the CVEs of interest.
3. Open the same project in the Safeguard web app and read the Reachability column.
4. Ask the platform chat for the call path behind any reachable verdict.
5. Record your decision, with a reason and an expiry date, for anything you defer.

Quick check

Sources

• Safeguard docs, Vulnerability prioritization and reachability
• Safeguard docs, ChatGPT (OpenAI)
• Safeguard docs, MCP tool availability and gating
• CISA Known Exploited Vulnerabilities Catalog

Last reviewed: 18 October 2026.