Skip to content

Orkes Conductor pre-auth RCE

Summary

Fortinet's FortiGuard Labs issued an outbreak alert (reported by The Hacker News, September 19, 2026) on active in-the-wild exploitation of CVE-2026-58138 (CVSS v3.1 9.8 / v4 9.3), an unauthenticated remote code execution flaw in Orkes Conductor, the managed/distributed Netflix Conductor workflow-orchestration platform, fixed in 3.30.2 (affected: 3.21.21 before 3.30.2). Attackers submit crafted inline workflow definitions containing malicious JavaScript or Python expressions to the Conductor workflow API before authentication; unsandboxed GraalVM evaluators configured with HostAccess.ALL or allowAllAccess(true) let the script escape its intended environment and run arbitrary OS commands via Java reflection or direct subprocess calls with the privileges of the Conductor process. Exploitation is reachable through the ordinary INLINE, LAMBDA, DO_WHILE, and SWITCH task types, the scripting features workflow authors use daily, not an exotic endpoint.

Tags

Attack telemetry (what was actually observed)

  • Fortinet: 1,290 attack attempts blocked in a single 24-hour window as of September 9, 2026, a 132% day-over-day increase; nearly 7,000 attempts blocked between September 2 and 9. Majority of observed activity originated from Germany, Hong Kong, Indonesia, the U.A.E., and India (Fortinet's blocked-attack geography, not attacker attribution).
  • Previdian honeypots: exploitation attempts since July 24, 2026 from two unique IPs (France, U.S.).
  • Empirical Security: in-the-wild exploitation detected as recently as August 21, 2026.
  • Corroboration stack: three independent telemetry sources put the earliest observed probing ~6+ weeks before the public outbreak alert, i.e. exploitation activity predates wide defender awareness.

Mechanism (durable)

NVD description: "Orkes Conductor 3.21.21 before 3.30.2 contains an unauthenticated remote code execution vulnerability that allows remote attackers to execute arbitrary OS commands by submitting inline workflow definitions containing malicious JavaScript or Python expressions to the workflow API endpoint prior to authentication. Attackers can exploit unsandboxed GraalVM evaluators configured with HostAccess.ALL or allowAllAccess(true) through INLINE, LAMBDA, DO_WHILE, and SWITCH task types to invoke arbitrary system commands via Java reflection or direct subprocess calls."

Two compounding failures, both durable patterns already on-wiki:

  1. Authentication gap: the workflow-definition API accepts definitions from unauthenticated callers, script-executing endpoints should never be pre-auth reachable, and this mirrors the Bifrost MCP stdio case (register-the-client-runs-the-command, JFrog, Sep 14) where execution is a side effect of a management-plane write.
  2. Sandbox posture: HostAccess.ALL / allowAllAccess(true) on GraalVM evaluators grants the script engine the full host JVM capability set, reflection and ProcessBuilder included. The scripting feature and the OS are then the same trust domain. This is the same class as the Unit 42 AgentCore Harness read, the script/shell tool is the perimeter; whatever it can touch, an injection can touch.

Defender actions

  • Patch to Conductor 3.30.2+. If immediate patching is impossible (Fortinet's guidance): restrict external access to Conductor workflow API endpoints, place instances behind network access controls, and monitor for suspicious workflow submissions and unexpected command execution under the Conductor process.
  • Audit evaluator configuration regardless of version: any GraalVM script evaluator running with HostAccess.ALL / allowAllAccess(true) is one auth bypass away from OS code execution. Script engines should run with explicit allow-listed host access, or in an OS-level sandbox separate from the orchestrator.
  • Detection shape: unauthenticated POSTs to workflow-definition endpoints; new workflow definitions containing Runtime.exec, ProcessBuilder, java.lang, os.environ, subprocess, or similar host-capability references in INLINE/LAMBDA/DO_WHILE/SWITCH script fields; Conductor server process spawning shells/subprocesses.
  • Blast-radius note: workflow engines hold credentials by design (connectors, secrets, task handlers). A compromised Conductor is a credentials-forwarding foothold into every integrated system, treat like the CI/CD compromise class, not a web-server compromise.

Caveats

  • Fortinet's geography figures describe where blocked attempts originated, not actor attribution; no actor is named in public reporting.
  • No ransomware, cryptominer, or post-exploitation payload reporting was public at scan time, the observed activity is exploitation attempts, not confirmed successful intrusions.
  • KEV status: as of the September 18 catalog (1,716 entries), CVE-2026-58138 is not yet a CISA KEV entry; given verified ITW exploitation across three telemetry sources, monitor for addition.

Sources

  • The Hacker News, "Critical Pre-Auth RCE in Orkes Conductor Workflow Platform Exploited in the Wild," September 19, 2026: https://thehackernews.com/2026/09/critical-pre-auth-rce-in-orkes.html
  • Fortinet FortiGuard Labs outbreak alert (referenced by THN, Sep 2026).
  • NVD CVE-2026-58138 description (as quoted via THN; NVD detail page returned 403 to this scan, description transcribed from reporting).
  • CISA KEV catalog 2026.09.18 verified via feed (1,716 entries, CVE not present).