Microsoft just gave every Entra ID administrator a hard deadline: migrate off SMS first-factor authentication by February 2027 or watch users get locked out. The same week, AI agents demonstrated - again - that identity controls for non-human entities remain dangerously immature. Google’s Gemini broke into three real companies during a scoped security test, researchers used Claude Opus 5 to chain identity flaws and take over OpenAI staff accounts, and a CVSS 9.8 pre-auth RCE in a workflow orchestration platform is being exploited in the wild - exposing every secret those workflows touch.

The through-line is clear: identity is the control plane, and the things authenticating to it are no longer just humans.

In the News

Microsoft Retiring Entra ID SMS First-Factor by February 2027

Microsoft has issued formal guidance to Entra ID administrators: SMS as a primary authentication factor will be retired by February 2027. After that date, users still enrolled with SMS-only authentication will face sign-in disruptions - not a soft deprecation, but a hard cutoff.

The directive pushes organizations toward passkeys and FIDO2 security keys as Microsoft’s preferred phishing-resistant path. This is not a surprise - Microsoft has been signaling this move for over a year - but the firm deadline turns it from a roadmap item into a project with a due date.

The operational challenge is inventory. Most Entra tenants have a mix of authentication methods - some users on passwordless, some on authenticator push, and a long tail still on SMS. Admins need to identify that long tail now, not in January 2027. The migration involves enrolling users in passkeys or FIDO2 keys, updating conditional access policies, and communicating the change to end users who are accustomed to receiving a text message.

What defenders should do: Run an Entra ID authentication methods report today to identify all users still registered with SMS as their primary factor. Begin enrollment in phishing-resistant methods - FIDO2 security keys or platform passkeys - and update conditional access policies to block SMS-only authentication before the February 2027 deadline.

Google’s Gemini AI Breached Three Real Companies During Security Test

Google disclosed that during a May 2026 security evaluation, its Gemini AI agent accessed production systems at three real companies it was never authorized to reach. The root cause was a domain mix-up during the test setup - the agent was scoped to target evaluation infrastructure but instead operated against live environments.

Google and the affected companies confirmed no data was exfiltrated, and the incident was contained once discovered. But the implications go beyond this specific failure. The Gemini incident follows a pattern: AI agents operating with real credentials, hitting real systems, and exceeding their intended scope. When the identity boundary for an AI agent is misconfigured, the agent does exactly what it was designed to do - execute against whatever target it can reach.

This is compounded by a separate research disclosure this week. Researchers from Hacktron used Anthropic’s Claude Opus 5 to autonomously discover and chain two identity flaws - a help-forum vulnerability and an authentication weakness - to take over several OpenAI employee accounts and reach internal code repositories. This was an authorized security research engagement, not a malicious breach. But it demonstrates that AI models can now act as autonomous identity-attack tools, chaining flaws without human guidance.

What defenders should do: Before deploying any AI agent, document what identity it uses, what access that identity has, and what blast radius exists if the agent operates outside its intended scope. Treat AI agent identities as non-human identities requiring the same lifecycle governance as service accounts - scoped access, time-bound credentials, and behavioral monitoring.

Orkes Conductor Pre-Auth RCE Exploited in the Wild - CVE-2026-58138

A pre-authentication remote code execution vulnerability in Orkes Conductor (CVE-2026-58138, CVSS 9.8) is being actively exploited. Fortinet confirmed wild exploitation. Conductor is a workflow orchestration platform used to automate business processes - and those workflows routinely carry service-account credentials, API tokens, and secrets as part of their execution.

A pre-auth RCE on the orchestrator means an attacker who reaches the Conductor instance can execute code without any credentials, then harvest every secret flowing through the orchestrated workflows. This is non-human identity exposure at scale - a single compromised Conductor instance can yield credentials for dozens of downstream systems.

What defenders should do: Patch Orkes Conductor immediately. After patching, audit every service account and API token that has been used in Conductor workflows - assume they were exposed and rotate them. Apply network segmentation to ensure orchestration platforms are not reachable from untrusted network segments.

AWS Auto-Locks Exposed IAM Keys Found on GitHub

Unit 42 published a detailed analysis of how AWS detects and neutralizes IAM credentials leaked to public GitHub repositories. The mechanism: GitHub’s secret-scanning identifies exposed AWS keys, notifies AWS, and AWS automatically applies managed policies that restrict the compromised credential’s access within minutes.

The research identifies the specific CloudTrail events that signal automated lockdown - primarily the attachment of AWS-managed quarantine policies to the compromised IAM user. These are high-fidelity detection signals. If your SIEM is not alerting on managed-policy attachment events from the AWS service principal, you are missing automated containment actions that affect your running applications.

What defenders should do: Add CloudTrail alerting for AWS-managed quarantine policy attachment events. Validate that secrets management practices prevent IAM keys from being committed to repositories in the first place - pre-commit hooks, secrets scanning in CI/CD pipelines, and short-lived credentials via IAM Roles Anywhere or OIDC federation.

Defender Action Items

  • Entra ID SMS migration: Run authentication methods reports now. Begin FIDO2/passkey enrollment for all users still on SMS-only. Update conditional access policies to enforce phishing-resistant methods before the February 2027 deadline.
  • Orkes Conductor CVE-2026-58138: Patch immediately. Rotate all service-account credentials and API tokens used in Conductor workflows. Segment Conductor instances away from untrusted networks.
  • AI agent identity governance: Inventory all AI agent identities (Gemini, Copilot, Codex, internal agents). Document their access scope, credential type, and blast radius. Apply time-bound credentials and behavioral monitoring.
  • AWS credential exposure monitoring: Add CloudTrail correlation rules for managed quarantine policy attachment. Implement pre-commit secret scanning and short-lived credentials to prevent key leakage.
  • SolarWinds ARM CVE-2026-28326: Update Access Rights Manager to version 2026.3+. Audit for hard-coded key exposure.

Detection Queries

The following Splunk SPL query detects when AWS automatically attaches a quarantine managed policy to an IAM user - the signal that GitHub secret-scanning has triggered an automated credential lockdown:

index=aws sourcetype="aws:cloudtrail" eventName="AttachUserPolicy"
| where userIdentity.invokedBy="access-analyzer.amazonaws.com" OR userIdentity.invokedBy="support.amazonaws.com"
| where match(requestParameters.policyArn, "AWSCompromisedKeyQuarantine")
| stats count by requestParameters.userName, requestParameters.policyArn, sourceIPAddress, awsRegion
| sort - count

This fires when AWS’s automated system (not a human administrator) attaches the AWSCompromisedKeyQuarantine policy. False positive rate is near zero - this policy is only attached by AWS’s automated credential-exposure response. Any hit means a credential was found in a public repository and is now restricted.

References

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.