CVE-2026-100348

The defect ---------- Confirmed at HEAD (f0407d5, v0.144.0). openapi3/loader.go, inside resolveComponent's drill closure (around line 455), the case *SchemaRef branch for pathPart == "additionalProperties" dereferences c.Value.AdditionalProperties with no nil-check on c.Value: case *SchemaRef: if pathPart == "additionalProperties" { if ap := c.Value.AdditionalProperties.Has; ap != nil { // c.Value can be nil here cursor = *ap } else { cursor = c.Value.AdditionalProperties.Schema } attempted = true } When a $ref fragment drills through .../additionalProperties of a schema whose Value is still nil (a pure, not-yet-resolved $ref, or a self-reference resolved before its own Value is set), c.Value is nil and the .AdditionalProperties access panics. Input and attacker ------------------- A tiny crafted OpenAPI document, one schema, no external refs, no network: openapi: 3.0.3 info: {title: t, version: "1.0.0"} paths: {} components: schemas: A: $ref: '#/components/schemas/A/additionalProperties' Attacker model: any service or tool that loads an untrusted spec via the default openapi3.Loader (LoadFromData / LoadFromFile / LoadFromURI), e.g. a spec registry, an API-management platform ingesting user specs, a linter, or a converter. It crashes during LOAD (ResolveRefsIn), before Validate. Verification level ------------------ EXECUTED with negative controls (agent-run go test at HEAD f0407d5, including yesterday's security merge; the bug is still present). The full stack pins LoadFromData:196 -> ResolveRefsIn:274 -> resolveSchemaRef:1048 -> resolveComponent:485 -> drill:439 -> :455, propagating to main(); no recover() in loader.go/schema.go (grep + trace). Two negative controls stay clean: drilling additionalProperties of a resolved schema loads fine (the existing TestIssue376 happy path), and a normal recursive schema loads fine, so the nil Value is specifically the cause. I independently re-confirmed at HEAD that loader.go:455 dereferences c.Value.AdditionalProperties with no nil-check on c.Value. NOT verified: I did not enumerate every consumer's recover posture. Suggested fix ------------- Add if c.Value == nil { break } (fall through to the nil-safe drillIntoField) before touching c.Value.AdditionalProperties. One line. Also cleared (stated, not claimed) ---------------------------------- The allOf/anyOf $ref diamond does not blow up: Schema.validate threads its stack through the return and never pops it, acting as a global pointer-keyed visited-set (executed N=10..30, microseconds). External-ref SSRF/path-traversal is blocked by default (IsExternalRefsAllowed=false); the known custom-ReadFromURIFunc bypass is the already-banked downstream oasdiff item. Advisory-source note -------------------- Five-source enumeration: every kin-openapi advisory (CVE-2025-30153 plus the five fresh July-2026 GHSAs) is in openapi3filter (request validation); this finding is in the openapi3 loader, untouched by that systematic reporter and flagged un-analyzed by the prior oasdiff scout. Not contested on the loader surface as of today (a systematic hunter could pivot, so it decays; submit reasonably promptly). Dedup ----- Dedup gate clean: no kin-openapi loader verdict in COVERED-INDEX, DISCLOSURE-INDEX or SCOUT-EXCLUDE (the one DISCLOSURE-INDEX hit is report #313, the downstream oasdiff, a different repo/lane). Tooling ------- AI assistance was used while investigating. The panic is executed-verified with negative controls (a one-schema self-referential additionalProperties $ref crashes LoadFromData while a resolved additionalProperties and a normal recursive schema load fine), and I confirmed at HEAD that loader.go:455 dereferences c.Value.AdditionalProperties with no nil-check on c.Value, on a load path with no recover.