Practical guide · November 17, 2026
How to validate AI-generated security findings
Treat an AI security report as a hypothesis until its attacker assumptions, execution path and security consequence are established in a relevant environment. A working reproducer is valuable evidence, but it does not by itself establish that an attacker can supply the input, reach the operation or obtain the impact claimed in the report.
The goal is not to distrust reports because AI helped write them. It is to apply the same evidence standard to every reporter while keeping the review workload manageable. This guide proposes a practical acceptance procedure for application-security teams and maintainers; it is not a claim that we have tested a particular discovery product.
Normalize the claim before reading the whole narrative
Ask for a short statement containing five elements: affected revision, attacker starting privileges, controlled input, boundary crossed and observed outcome. Keep proposed impact separate from demonstrated impact. A report saying “critical authorization bypass” is less useful than one explaining which identity accessed which prohibited resource and how that result was checked.
This matters even when reports improve. In an October 22, 2026 account, curl maintainer Daniel Stenberg says hallucinated reports had largely disappeared in that project's experience, but report volume, verbosity, bug-versus-vulnerability judgments and severity assessment remained problems. That is one maintainer's account, not a measured false-positive rate for AI tools generally.
Our recommendation is to accept a concise claim first, then inspect the supporting material. Do not reject a real security failure because the explanation is verbose; make the evidence easier to review. Conversely, polished prose should not compensate for a missing precondition.
Trace attacker control across every boundary
For a proposed source-to-sink issue, work backward from the unsafe operation. Identify each transformation, validation step and permission check. Then work forward from the attacker's entry point. The two paths must meet under compatible conditions.
Ask whether the input is an external request, an authenticated user's field, an administrator-selected setting or a local test fixture. If reproduction requires an administrator to disable a control, describe that assumption. It might still reveal an important defense-in-depth issue, but it is not automatically an unauthenticated attack.
OWASP's authorization-testing guidance distinguishes access across users and roles. We extend that distinction into a review rule: preserve the actor, resource owner and expected permission in every finding, rather than treating a missing check in one function as sufficient proof.
Hypothetical example: a tenant-export report
Suppose a report flags an export handler because its query does not explicitly filter by tenant. Your service is intended to allow an ordinary user to export only their own tenant's records. Use synthetic tenants Cedar and Maple in an authorized test environment.
First establish the positive control: a Cedar user can export Cedar data. Then test the disputed boundary: the same Cedar identity must not obtain Maple data. Record the actual authenticated identity and the object's ownership independently of whatever the request claims. Also inspect whether middleware, a shared data-access layer or database policy already enforces the boundary.
If Maple data becomes accessible only after a test fixture bypasses the service's identity layer, the fixture has not demonstrated the original claim. If the normal service returns Maple data to the Cedar identity, the evidence is much stronger. If the environment cannot start, label the result unresolved; do not convert a plausible code path into a reproduced cross-tenant disclosure.
Reproduce without replacing the application you are assessing
Preserve the vulnerable revision, dependency lockfiles, build configuration and minimal input. Record any changes required to run the test. Instrumentation that records behavior is different from a patch that removes a permission check or directly calls an internal function with impossible state.
Where safe, start from a clean environment and have a reviewer replay the evidence without relying on the generating agent's explanation. Store the command or request sequence, relevant output, environment identifiers and cleanup steps privately. Exclude secrets and customer records from report copies.
A failure to reproduce has several possible dispositions: incorrect claim, missing setup information, unsupported configuration, nondeterministic failure or unavailable environment. Keep those separate. “Could not reproduce” is not equivalent to “proved safe,” just as “agent reported success” is not equivalent to independently observed success.
For unsafe or impractical reproduction, accept a clearly labeled static finding when the evidence justifies it. Specify which predicates remain unverified and who owns the next check. Do not demand a destructive demonstration merely to satisfy a reproducibility checkbox.
Separate the symptom from the security consequence
Our acceptance procedure asks for the narrowest outcome the evidence establishes. A crash demonstrates a failure under the recorded conditions, not automatically code execution. A response containing an object identifier does not necessarily expose protected content. A timing difference is not automatically a practical secret-recovery method.
Write impact as a chain of supported statements: the actor can cause this operation; the operation violates this expected boundary; the observed consequence affects this asset. If a further step is hypothetical, label it as such. Severity should follow the established conditions, not the headline generated before validation.
Use distinct queue states: confirmed security finding, confirmed nonsecurity defect, hardening proposal, duplicate root cause, unresolved candidate and rejected claim with reason. These are proposed workflow labels. Keep unresolved candidates visible rather than hiding them inside a favorable precision score.
Test the patch against the original failure and legitimate use
For a reproduced finding, retain a regression test that fails on the vulnerable version and passes with the fix. Add a positive control for intended behavior. In the tenant example, the patch should deny the prohibited export while still permitting the user's own export. Disabling exports altogether may suppress the demonstration without meeting the product requirement.
Review whether the patch addresses the shared cause or only the example. Identify related callers and alternate entry points explicitly; do not claim they were tested unless they were. A reviewer other than the generating agent should inspect consequential changes. OpenAI's documented defender workflow likewise describes engineer review and regression verification where feasible; that documentation does not establish success for any particular patch.
Measure a pilot without hiding the difficult cases
Before running a product comparison, define the evaluation unit as a distinct root cause with a documented security boundary. A hundred generated inputs that trigger the same defect should not become a hundred discovery wins. Use the same repository revision and agreed scope for the baseline and candidate, and record differences in access, configuration and time budget.
Our recommended pilot report includes adjudicated precision, unresolved count, unique new confirmed findings, duplicate count, reviewer time and setup failures. Define adjudicated precision as confirmed security root causes divided by all adjudicated security claims after deduplication: confirmed, rejected, and claims that proved to be nonsecurity defects. Also report the nonsecurity-defect count and the unresolved count separately, so readers can see both useful defect yield and security-classification errors. This is a local operational definition, not a universal benchmark metric.
Include reviewer time spent disproving claims as well as time spent confirming them. Report engineering time to validate patches separately from discovery time. A tool that finds useful defects but exhausts the only available reviewer creates a different operational result from a self-sufficient report pipeline.
Do not claim recall for an ordinary private repository when the complete set of vulnerabilities is unknown. With a held-out set of known defects, report recovery on that set, how it was selected, and whether the system saw advisories, fixes or vulnerability descriptions. That result is not the fraction of all vulnerabilities in production that the tool would detect.
Read benchmark tasks literally
CyberGym's repository describes a benchmark built around real-world vulnerabilities and executable validation tasks. Inspect the specific task's supplied artifacts and success condition before translating a benchmark score into a buying claim. Generating a reproducer with an existing vulnerability description is a different task from discovering an unknown flaw with no such description.
For your own evaluation, separate discovery, reproduction and repair tracks. State what information each track receives. Keep held-out fixes and answer artifacts inaccessible to the discovery track, and record model/configuration changes between runs. These are proposed controls; we have not audited any vendor's benchmark contamination or independently reproduced its score.
Once the finding is established, move to remediation prioritization. If an external provider is delivering the work, include the evidence requirements in the penetration-test report agreement. The acceptance standard is a defensible claim with bounded uncertainty—not a report that happens to sound certain.
Evidence reviewed November 17, 2026. AI-assisted, documentation-led editorial analysis; no product or benchmark runs were performed. The tenant example is fictional. Validation states, controls and pilot metrics are proposed methods, not measured results.