Three stories dominate the identity threat surface this morning: Microsoft proved that device code phishing - a technique defenders already track against human users - extends cleanly to AI coding agents. VMware shipped a patch for a CVSS 9.8 auth bypass that Chinese espionage operators have been exploiting for two months. And the FBI quantified what happens when AI agents hunt for leaked credentials at machine speed.
The common thread is that identity is the attack surface, whether the victim is a human, an AI agent, or a service account that nobody remembered to rotate.
In the News
Device Code Phishing Now Targets AI Coding Agents via OAuth Hijacking
Microsoft published research demonstrating that the OAuth device code flow - already a known phishing vector against human users - works against AI coding agents with no modification. Using GitHub Copilot CLI as a proof-of-concept target, researchers showed that an attacker who tricks a user into entering a device code during an agent authentication flow gains a persistent session. That session inherits the user’s full identity scope: repository access, command execution, and cross-session persistence.
The mechanism is straightforward. Device code authentication asks users to visit a URL and enter a short alphanumeric code. The flow was designed for input-constrained devices (smart TVs, IoT). AI agents adopted it because they operate in terminal environments where browser-based OAuth redirects are impractical. The problem is that the code itself carries no binding to the legitimate session - if an attacker presents their own code, the user authenticates the attacker’s session instead.
This is not theoretical. Device code phishing against human targets is a documented technique (MITRE ATT&CK T1528). What changes is the blast radius: an AI agent’s session often has broader persistent access than a single-tab browser session, and agents may not trigger the same conditional access policies that protect human logins.
What defenders should do: Restrict or disable device code flow in conditional access policies for any tenant where AI agents authenticate. Enforce phishing-resistant passwordless authentication (FIDO2) for the humans who authorize agent sessions. Deploy identity threat detection that flags device-code grants from unexpected IP ranges or devices.
VMware vCenter Auth Bypass Exploited by China-Linked Group Since September
Broadcom patched CVE-2026-123890 on November 20 - a CVSS 9.8 authentication bypass in vCenter Server’s Directory Service. The flaw allows unauthenticated remote attackers to gain full administrative access to the vCenter control plane. BleepingComputer reports that China-linked threat actor Armory Castle has exploited it since at least September 2026.
vCenter is the management plane for VMware virtual infrastructure. Admin access to vCenter means control over every virtual machine, every host, every network segment the platform manages. The Directory Service component is vCenter’s identity backbone - it handles authentication and authorization for the entire environment. An auth bypass here is not a peripheral vulnerability; it is a complete identity compromise of the control plane.
Two months of dwell time is the baseline assumption for any organization that was running an unpatched vCenter instance accessible from Armory Castle’s operational infrastructure.
What defenders should do: Patch immediately. Audit vCenter Directory Service authentication logs for anomalous admin sessions dating back to September. Restrict management plane access via network segmentation - vCenter should never be reachable from general-purpose network segments. Deploy ITDR that monitors for admin-level identity creation or privilege escalation on infrastructure management platforms.
FBI Alert - ARTEX AI Agents Automate MSP Supply-Chain Attacks
The FBI issued a FLASH alert on November 13 detailing how ARTEX-powered AI agents automated a ConnectWise supply-chain intrusion. The agents moved from initial targeting to privileged access in under three hours, then deployed SCKit worms into MSP customer Microsoft 365 environments. The speed was enabled by non-human identity abuse: the agents systematically located and exploited leaked credentials, API keys, and service account tokens.
GitGuardian research underpinning the alert provides the scale context. An ARTEX build located 8 of 9 intentionally planted credentials in 6 to 13 seconds - without using a dedicated secrets scanner. Across the broader dataset, 64% of secrets leaked in 2022 were still valid in 2026. Storm-3358, a threat actor using ARTEX tooling, reached Azure Key Vault in under 5 minutes via public GitHub searches for exposed credentials.
This is what non-human identity abuse looks like at machine speed. The credentials are years old. Nobody rotated them. An AI agent finds them faster than any human analyst.
What defenders should do: Audit and rotate every service account credential, API key, and machine identity in the environment. Implement automated secrets rotation with a maximum lifetime policy. Deploy NHI governance tooling that inventories every non-human identity and flags credentials that have not been rotated within policy. Monitor for credential use from unexpected sources - ITDR is the detection layer here.
COLD RELIC Pivots Across US Organizations Using Stolen Google Workspace Tokens
Russian espionage group COLD RELIC used credentials and session tokens stolen from one Google Workspace breach to gain initial access to a second US organization. No malware was required. No software vulnerability was exploited. The attack was pure identity-based lateral movement - stolen cloud IAM artifacts served as the bridge across organizational trust boundaries.
This pattern is increasingly common in state-sponsored operations. Cloud identity is the perimeter. When a session token from Organization A works in Organization B’s environment (via shared SaaS platforms, federated trust, or OAuth grants), the blast radius of a single identity compromise extends across the supply chain.
What defenders should do: Enforce device-bound session tokens and conditional access policies that validate device posture on every authentication. Monitor for sessions originating from unexpected geolocations, ISPs, or device fingerprints. Review OAuth app grants for overly permissive cross-tenant access.
Defender Action Items
- Patch CVE-2026-123890 (vCenter) immediately - this is actively exploited with two months of potential dwell time. Audit Directory Service logs back to September.
- Patch CVE-2026-114938 (Java4Shell) on all identity-critical Java services - Keycloak, Jira, Jenkins are confirmed exploitable. CISA KEV-listed November 18.
- Patch CVE-2026-115731 (Puppet) - compromised Puppet servers expose every secret and access policy they provision. CISA KEV-listed November 18.
- Restrict OAuth device code flow in conditional access policies for tenants where AI agents authenticate.
- Rotate all NHI credentials (API keys, service accounts, machine identities) and enforce maximum-lifetime rotation policies.
- Enforce phishing-resistant passwordless authentication (FIDO2) for all admin and privileged access - especially on infrastructure management planes.
Detection Queries
The following Splunk SPL query surfaces anomalous device-code OAuth grants - the mechanism Microsoft demonstrated against AI coding agents. Look for device-code grants from unexpected IP ranges or for service principals that should not be using the device code flow.
index=azure_ad sourcetype="azure:aad:signin"
authentication_protocol="deviceCode"
| eval grant_time=strftime(_time, "%Y-%m-%d %H:%M:%S")
| stats count AS device_code_grants, values(src_ip) AS source_ips, values(app_display_name) AS apps, latest(grant_time) AS last_grant BY user_principal_name
| where device_code_grants > 1
| sort -device_code_grants
This query isolates device-code authentication events in Entra ID sign-in logs and aggregates them by user principal. Any user with multiple device-code grants - or grants to unexpected applications - warrants investigation. False positive rate is low in environments where device-code flow is restricted by policy; high in environments that have not restricted it.
References
- Microsoft - The risks of device code authentication for AI coding agents - Microsoft Security Blog
- VMware patches vCenter zero-day exploited in attacks since September - BleepingComputer
- FBI FLASH: ARTEX AI agents automate ConnectWise supply-chain attack - GitGuardian Blog
- Russian COLD RELIC hackers use stolen Google Workspace data to pivot - The Hacker News
- Java4Shell CVE-2026-114938 enables pre-auth RCE - The Hacker News
- Hackers exploiting critical Puppet RCE - The Hacker News
- React2STEAL: hijacked GitHub accounts exploit React2Shell - GitGuardian Blog
- GitGuardian: financial services secrets exposure and ARTEX - GitGuardian Blog
- Microsoft - Security for the agentic era - Microsoft Security Blog
- Okta - NGINX2Root security updates - Okta Security
Related Briefs
- vCenter Auth Bypass Zero-Day - CVE-2026-123890
- NGINX HTTP/2 Pre-Auth RCE - CVE-2026-123007
- AI Agents Steal Keys to Find More Keys - NHI Sprawl
- Java4Shell CVE-2026-114938 - Pre-Auth RCE Hits Jira
- Optimizely Supply-Chain Attack Hits 178 Orgs
Subscribe to The Identity Brief
Get The Identity Brief in your inbox (Mon/Wed/Fri) - Human, machine, and AI identity security — NHI, ITDR, and the IAM market.