The Intake — Saturday, September 12, 2026

If your agent pipeline runs Orkes Conductor and that instance is reachable from any network, an unauthenticated attacker can execute arbitrary code on that host. Orkes Conductor is a workflow orchestration engine used in AI agent pipelines.

The community API endpoint evaluates user-supplied JavaScript and Python expressions. The evaluation runs through a GraalVM scripting context. That context is configured with HostAccess.ALL, giving the scripting layer full access to the host system. No authentication is required on the endpoint by default. An attacker with network access executes code as the Conductor process user. The flaw affects versions 3.21.21 through 3.30.1. CVE-2026-58138, CVSS 9.8. Active exploitation has been confirmed since late August 2026. A public proof-of-concept is available on GitHub and a corresponding Exploit-DB entry (52633) is published. Empirical Security reports the fix is version 3.30.2.

If you're on any version from 3.21.21 through 3.30.1 with the community API endpoint accessible from a network, 3.30.2 is the patched version.

If you're running Google Cloud Agent Development Kit (ADK) for Python 2.0.0 through 2.6.0 in any environment where pytest is installed, the environment has an unauthenticated code execution path.

Google Cloud ADK for Python is the toolkit for building agents on Google Cloud. The vulnerability is in the test session replay mechanism. A crafted replay file lets an unauthenticated remote attacker execute arbitrary code. Pytest must be present in the same environment for the attack path to exist. Cloud Run, GKE, and open-source deployments with pytest all qualify. The flaw is CVE-2026-79696, published September 9, 2026. CVSS 4.0 scores it 10.0 Critical. No workaround short of upgrading is documented. The fix is ADK version 2.7.0.

If you're on ADK Python 2.0.0 through 2.6.0 and pytest is in your environment, 2.7.0 is the patch.

If you're building agents that initiate payments or financial commitments through Visa, Mastercard, or Ant network infrastructure, the authentication layer those transactions will need is starting to take shape.

Ant International, Mastercard, and Visa announced September 9–10 that they are jointly developing a KYA interoperability standard. The standard merges three existing frameworks: Visa's Trusted Agent Protocol, Mastercard Verifiable Intent, and Ant's Agentic Mobile Protocol. Governance runs through BuildFin.ai, a standards body convened by the Monetary Authority of Singapore. The parties project AI agents will handle $3–5 trillion in global commerce by 2030. No methodology for that projection was published. No compliance requirement or adoption timeline has been announced.

If you're building payment-initiating agents, the shape of the identity problem is now named. Three separate frameworks are converging, but no specification exists yet to build against.

---