GraphQLer: Enhancing GraphQL Security with Context-Aware API Testing

Omar Tsai, Jianing Li, Tsz Tung Cheung, Lejing Huang, Hao Zhu, Jianrui Xiao, Iman Sharafaldin, Mohammad A. Tayebi · 2025

Read GraphQLer: Enhancing GraphQL Security with Context-Aware API Testing by Omar Tsai, Jianing Li, Tsz Tung Cheung, Lejing Huang, Hao Zhu, Jianrui Xiao, Iman Sharafaldin, Mohammad A. Tayebi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-13358.txt.

Introduction

Omar Tsai and co-authors present a tool that hunts security flaws in GraphQL interfaces, the query systems behind many finance, shop, and social apps. Their 2025 paper notes that the worst flaws need chained steps, such as make, delete, then read, which standard scanners miss since they test each call alone.

The tool maps how calls depend on each other, then builds chains aimed at known flaw kinds. On a live finance system it flagged eight possible flaws with no login or docs in hand. On a shop system with known bugs it replayed all four cited access flaws. Across public systems its coverage far beat two rival tools, and it found chained flaws the rivals missed entirely.

Catalog record

CREATOROmar Tsai, Jianing Li, Tsz Tung Cheung, Lejing Huang, Hao Zhu, Jianrui Xiao, Iman Sharafaldin, Mohammad A. Tayebi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-13358

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

GraphQLer: Enhancing GraphQL Security with Context-Aware API Testing Omar Tsai, Jianing Li, Tsz Tung Cheung, Lejing Huang, Hao Zhu, Jianrui Xiao, Iman Sharafaldin, Mohammad A. Tayebi Abstract . GraphQL APIs power production systems across financial services, e-commerce, and social platforms, yet their most critical access-control vulnerabilities—Insecure Direct Object Reference (IDOR), Use-After-Free (UAF), and state-dependent injection—routinely escape automated security testing. Industry-standard scanners ( ZAP ) and the leading research fuzzer ( EvoMaster ) test operations in isolation and cannot compose the multi-step sequences these flaws require. We present GraphQLer , an open-source automated security testing framework built for production GraphQL APIs. GraphQLer constructs a typed dependency graph from live schema introspection and synthesizes vulnerability chains —ordered operation sequences targeting specific flaw classes. Three strategies cover the critical attack surface: topological SCC-traversal for general reachability, cross-user IDOR replay for broken access control, and CREATE → \to DELETE → \to READ synthesis for UAF. On a production financial API ( FinServ ), GraphQLer identified eight potential vulnerabilities—including denial-of-service vectors that exposed stack traces and sensitive implementation details—without prior documentation or authentication credentials. On a self-hosted Saleor instance pinned to the CVE-2022-39275 commit, GraphQLer reproduced all four broken-access-control mutations cited in the security advisory. On the 11 public APIs of the coverage set, GraphQLer achieves 85.52% mean PositiveCoverage versus 29.29% ( EvoMaster ) and 21.80% ( ZAP ); across the 21 evaluated APIs it detects all 5 confirmed IDOR endpoints, UAF behavior on two controlled schemas, and confirms XSS and SQLi on DVGA (an independent third-party oracle)—while all baselines detect zero chain-based vulnerabilities. \correspondingauthor DOI: 10.1145/3832783.3834496 ISBN: 979-8-4007-2882-2/2026/10 Conference: Proceedings of the 41st IEEE/ACM International Conference on Automated Software Engineering; October 12–16, 2026; Munich, Germany Proceedings of the 41st IEEE/ACM International Conference on Automated Software Engineering (ASE ’26), October 12–16, 2026, Munich, Germany ase26ind-p143-p CCS: Security and privacy Web application security CCS: Security and privacy Software security engineering CCS: Security and privacy Vulnerability scanners CCS: Software and its engineering Software testing and debugging CCS: Information systems Query languages † † cc-license: by 1. Introduction GraphQL has emerged as a dominant API technology, adopted widely for its schema-driven, single-endpoint design that lets clients compose arbitrary queries over richly typed data graphs ( Quiña-Mera et al., 2023 ; Brito and Valente, 2020 ) . This expressiveness, however, introduces security challenges largely absent from REST APIs. Unlike REST, where HTTP verbs and URL structure encode intent, GraphQL exposes all operations through a single /graphql endpoint with no built-in access-control enforcement. Mutations can silently depend on resources created by earlier operations, and the server resolves these dependencies at runtime—invisible to any tool that tests operations in isolation. This stateful dependency structure gives rise to three practically important vulnerability classes. Insecure Direct Object Reference (IDOR) : a resource created by one user can be read or modified by a second, unprivileged user who supplies only the resource ID. Use-After-Free (UAF) : a soft-deleted resource remains accessible after deletion because the backing store is never purged—the same free-then-dereference pattern as memory-level UAF, manifesting here as unauthorized data disclosure. Injection : SQL, OS command, and XSS payloads may require API state initialized by prior mutations—particularly on authenticated or multi-tenant APIs where valid resource IDs gate access. IDOR and UAF share a structural property that makes them invisible to existing tools: no single operation, executed in isolation, is sufficient to trigger the vulnerability . Injection occupies a middle ground: payloads can sometimes be sent in isolation on simple APIs, but in practice—on authenticated or multi-tenant APIs—exploiting injection requires valid resource IDs materialized by prior mutations, state that single-operation tools never construct. Existing tools fail to model this stateful structure. Black-box scanners (ZAP ( OWASP, ) , BurpSuite ( PortSwigger, ) ) vary parameters of individual operations and flag HTTP-level anomalies; they never compose multi-step sequences. EvoMaster ( Belhadi et al., 2023 ) —the only published GraphQL fuzzer we reproduce for comparison—applies evolutionary search over payloads but does not synthesize cross-user or post-delete chains. Other published GraphQL testing work, including reinforcement-learning ( Saito et al., 2025 ) and LLM-assisted ( Liu et al., 2025 ) approaches, likewise adapts payloads for individual operations rather than composing sequences. Static analyzers ( Lambers et al., 2024 ; Karlsson et al., 2021 ; Vargas et al., 2018 ) detect structural anti-patterns but miss runtime access-control flaws. As a result, IDOR and UAF vulnerabilities in GraphQL APIs routinely escape automated testing. We present GraphQLer , an automated security testing framework that, to our knowledge, is the first to treat vulnerability chains —ordered operation sequences designed to expose a specific flaw class—as first-class constructs in GraphQL testing. GraphQLer operates in two phases. The compilation phase parses the introspection schema and constructs a typed dependency graph linking objects, queries, and mutations. The testing phase traverses this graph using three specialized chain strategies that synthesize and execute the sequences needed to expose IDOR, UAF, and injection vulnerabilities. An optional LLM-assisted resolver handles ambiguous mutation semantics; a multi-auth profile engine manages credentials for cross-user scenarios. GraphQLer also exposes its full pipeline as a Model Context Protocol ( Anthropic, 2024 ) server—a tooling contribution enabling AI-assisted security workflows. Industry Impact. On the production-scale FinServ API, GraphQLer operated fully black-box and identified eight potential vulnerabilities, including at least two denial-of-service vectors exposing stack traces and implementation details. The findings were shared with the security team, demonstrating an automated evidence-collection workflow for industrial triage. We make the following contributions: \diamond We introduce vulnerability chains —ordered GraphQL operation sequences—and three chain strategies (topological SCC-condensation, IDOR cross-user replay, UAF graph-edge synthesis) that detect multi-step vulnerabilities missed by every existing tool, including production systems carrying real security debt. \diamond We present GraphQLer , an open-source tool 1 1 1 Source: https://github.com/omar2535/GraphQLer . implementing all three strategies together with an injection suite covering nine vulnerability categories (three confirmed on purpose-built ground-truth APIs), an LLM-assisted dependency resolver, and a multi-auth profile engine. \diamond We evaluate GraphQLer on 21 APIs (11 public, 5 platform, 2 known-vulnerable, 3 purpose-built ground-truth). On the 11 public APIs of the coverage set, GraphQLer achieves 85.52% mean PositiveCoverage , versus 29.29% for EvoMaster and 21.80% for ZAP ; on the ground-truth APIs it detects all 5 confirmed IDOR endpoints, UAF behavior on two controlled schemas, and three injection categories; and it independently validates injection detection on DVGA. 2. Background & Related Work 2.1. GraphQL Fundamentals GraphQL ( Quiña-Mera et al., 2023 ) is a strongly-typed API language where queries, mutations, and subscriptions are exposed through a single /graphql endpoint. Introspection returns schema metadata at runtime and is GraphQLer ’s compilation input. Unlike REST, mutation intent is not encoded by HTTP verbs, so create/update/delete semantics must be inferred from names and descriptions. GraphQLer uses this schema view to infer hard/soft dependencies and synthesize executable chains. 2.2. Related Work GraphQL Testing. Existing GraphQL tools operate on single operations. Vargas et al. ( Vargas et al., 2018 ) and Karlsson et al. ( Karlsson et al., 2021 ) generate queries from the schema. Zetterlund et al. ( Zetterlund et al., 2022 ) replay production traces (white-box). Belhadi et al. ( Belhadi et al., 2023 ) extend EvoMaster ( Arcuri, 2018 ; Arcuri et al., 2021 ) with GraphQL plugins;

Read the complete text