Skip to content
eastbaycyber

Cybersecurity Threat Digest November 17, 2026: Microsoft 365 Token Theft and Critical AI Infrastructure Flaws

Threat digests 10 min read
EC
East Bay Cyber Editorial Team Updated
Week of 17 NOV 2026

TL;DR - Phishing kits are stealing Microsoft 365 session tokens that can bypass password and MFA protections. - Patch exposed AI, sandbox, CMS, and Open Object Rexx systems, and investigate token reuse. - Treat suspected identity compromise and internet-facing critical flaws as urgent priorities.

This cybersecurity threat digest for November 17, 2026 covers Microsoft 365 token theft, major incident reports, defensive AI developments, and critical AI infrastructure vulnerabilities. For background, see our digest covering recent router flaws and PHP2Shell and glossary explanation of what DOS means.

Top Stories#

Modern phishing kits target Microsoft 365 sessions

Huntress researchers detailed phishing kits named EvilTokens, Knight Office, and Kali365 that capture and replay Microsoft 365 session tokens. Unlike password-only phishing, adversary-in-the-middle techniques can steal authenticated browser sessions and potentially bypass protections satisfied during the original login.

Read the BleepingComputer report on modern Microsoft 365 phishing kits and prioritize these response actions:

  • Review identity-provider sign-ins for unfamiliar devices, locations, autonomous systems, and user agents.
  • Investigate refresh-token activity and session reuse after suspicious authentication events.
  • Revoke active sessions and refresh tokens for affected users.
  • Reset credentials after containment, while recognizing that password resets alone may not invalidate stolen sessions.
  • Remove unauthorized OAuth grants, mailbox rules, forwarding settings, and newly registered authentication methods.
  • Prefer phishing-resistant authentication, such as FIDO2 security keys or passkeys, where supported.

Ipsos reports a cybersecurity incident

The Register reported that opinion polling company Ipsos disclosed a cybersecurity incident. The available report does not establish the incident’s cause, scope, or affected data. Organizations should avoid assuming a specific intrusion method or impact.

Security teams that use Ipsos or related suppliers should:

  1. Check vendor communications and contractual notification channels.
  2. Confirm whether internal data, credentials, integrations, or hosted services are involved.
  3. Search for suspicious authentication and data-transfer activity associated with the relationship.
  4. Require validated indicators and remediation details before closing the review.

Source: The Register coverage via Google News.

Reported attack affects 178 major websites

Newsweek reported on a cyberattack affecting 178 major websites and listed organizations reportedly impacted. The supplied reporting does not independently verify the full list, attack vector, or operational impact.

Organizations named in the report should validate:

  • Whether their domains, applications, or APIs experienced availability or integrity issues.
  • Whether a common hosting, CDN, DNS, SaaS, or managed-service provider is involved.
  • Web-server, WAF, DNS, identity, and cloud audit logs for the relevant period.
  • Notifications from regulators, customers, partners, and incident-response providers.

Source: Newsweek coverage via Google News.

Defensive cybersecurity AI investment expands

SecurityWeek reported that Corma raised $125 million to train foundation models for defensive cybersecurity. Separately, Quartz reported that OpenAI is providing NATO access to its Daybreak cybersecurity AI models.

These announcements do not replace validated security controls. Before adopting AI-enabled defensive services, assess:

  • What telemetry and sensitive data leave the organization.
  • Whether prompts, logs, or model outputs are retained or reused.
  • Human approval requirements for containment and production changes.
  • Model access controls, tenant isolation, auditability, and abuse prevention.
  • Supply-chain, availability, and vendor-incident obligations.

Sources: SecurityWeek coverage via Google News and Quartz coverage via Google News.

Microsoft tests Remote Lock for Windows devices

Microsoft is testing Remote Lock, which would let enterprise administrators lock lost or stolen Windows devices without immediately wiping them. The feature could preserve evidence and reduce exposure while administrators determine whether recovery or erase actions are appropriate.

Source: BleepingComputer report on Windows Remote Lock.

OT security requires continuity-aware controls

Industrial Cyber examined why continuous-process manufacturing requires a different OT cybersecurity approach. Plants that cannot safely stop must balance security changes against safety, availability, and process stability.

Practical priorities include:

  • Map safety-critical and production-critical dependencies before changing controls.
  • Use passive monitoring where active scanning could disrupt fragile systems.
  • Test patches and compensating controls in representative environments.
  • Define emergency change procedures with operations, engineering, and safety teams.
  • Segment IT and OT while preserving explicitly documented operational flows.

Source: Industrial Cyber coverage via Google News.

University of Kansas announces cybersecurity research partnership

The University of Kansas announced a research partnership with a Kansas-based AI and cybersecurity company. The announcement concerns research and industry collaboration; the supplied information does not identify a security incident or product vulnerability.

Source: University of Kansas partnership coverage via Google News.

Analyst’s Take: Start with identity containment and patching the exposed critical components. Stolen Microsoft 365 sessions may survive a password reset, while the Phoenix, go1cms, and CubeSandbox flaws can expose secrets, alter records, or break tenant isolation. The incident reports involving Ipsos and 178 websites remain limited by the supplied reporting, so validate those claims through vendor and internal telemetry before drawing conclusions.

Critical Vulnerabilities#

CVE-2026-53998: Arize Phoenix server-side template injection

Arize Phoenix versions 13.0.0 through 15.6.0 contain an unauthenticated server-side template injection vulnerability in the applyChatTemplate GraphQL query. An attacker may traverse Python object internals to access environment variables and module state. Exposure of PHOENIX_SECRET could enable administrative session-token forgery and credential decryption.

Upgrade exposed deployments immediately. Restrict GraphQL and administrative interfaces, rotate PHOENIX_SECRET and other potentially exposed credentials, and review access logs for suspicious query activity.

CVE-2026-130071: go1cms unauthenticated document modification

TridotsTech go1cms contains a missing-authentication flaw in the guest-whitelisted update_doc method. Unauthenticated attackers can modify arbitrary documents, including User records, potentially enabling privilege escalation, account takeover, or business-data tampering.

  • Severity: CVSS 9.8
  • Fixed version: The supplied data identifies a fixing commit rather than a release version.
  • CISA KEV status: Not listed in the supplied data
  • References: Fixing commit and VulnCheck advisory

Apply the vendor-referenced fix, restrict administrative endpoints, and review User records, privilege assignments, and application data for unauthorized changes.

CVE-2026-120699: CubeSandbox host-path mounting

CubeSandbox versions before 0.5.1 allow an authenticated tenant to provide an arbitrary host path through metadata["host-mount"]. The control plane mounts that path as root, potentially enabling host-file access and modification, host code execution, container-runtime access, and compromise of co-resident tenants.

Upgrade control-plane and worker components as applicable. Audit tenant configurations, host mounts, container-runtime access, and cross-tenant boundaries. Treat the environment as potentially compromised if unauthorized host mounts were possible.

CVE-2026-130320: Open Object Rexx bounds-checking flaw

Open Object Rexx versions before 5.2.0 contain insufficient bounds checking in RXAPI nameArg and userid handling. Crafted messages may result in arbitrary code execution.

Upgrade to 5.2.0 and identify systems that expose RXAPI to untrusted local users, processes, or network sources.

CVE-2026-130327: Open Object Rexx local RXAPI injection

Open Object Rexx versions before 5.2.0 rely on a world-writable /tmp file or localhost TCP port 10010 for RXAPI communication. A local user may inject a macro that a REXX script executes.

Upgrade promptly. Until then, restrict local access, review processes using RXAPI, and assess whether /tmp permissions or port 10010 exposure could permit injection.

None of the listed vulnerabilities is marked as a CISA Known Exploited Vulnerability in the supplied data. That designation should not be treated as a safety signal: the Phoenix, go1cms, and CubeSandbox issues have especially significant impact because they involve unauthenticated access, arbitrary modification, secret exposure, or tenant-isolation failure.

What Defenders Should Do Today#

  1. Investigate Microsoft 365 token abuse. Search identity-provider, Microsoft 365, endpoint, and proxy telemetry for anomalous token reuse, unfamiliar refresh activity, impossible travel, new device registrations, suspicious OAuth consent, and sign-ins from unexpected infrastructure.

  2. Contain suspected identity compromise. Revoke active sessions and refresh tokens, reset credentials, remove malicious mailbox rules and OAuth grants, review authentication methods, and require phishing-resistant MFA where available.

  3. Patch the critical vulnerable components. Upgrade Arize Phoenix to 15.7.0, apply the go1cms fixing revision, upgrade CubeSandbox to 0.5.1, and upgrade Open Object Rexx to 5.2.0.

  4. Reduce exposure while patching. Remove public access to administrative interfaces and GraphQL endpoints, restrict RXAPI access, segment AI platforms and sandbox control planes, and use compensating controls for systems that cannot be upgraded immediately.

  5. Rotate exposed secrets. Review PHOENIX_SECRET, application credentials, service tokens, container credentials, and other secrets accessible from affected systems. Rotate them after containment and validate dependent services.

  6. Audit CubeSandbox isolation. Search configurations and audit logs for metadata["host-mount"], inspect host-file and container-runtime access, and validate tenant separation after remediation.

  7. Review incident and supplier exposure. Assess whether your organization, domains, applications, or suppliers appear in reporting about the attack affecting 178 websites or the Ipsos incident. Use verified incident notifications rather than third-party headlines alone.

  8. Test lost-device procedures. Prepare remote-lock workflows for Windows endpoints, document authorization and recovery steps, and preserve evidence before wiping devices when theft or compromise is suspected.

  9. Control defensive AI adoption. Define data-handling, access-control, human-approval, logging, model-risk, and vendor-incident requirements before connecting AI security tools to production telemetry or response systems.

Technical Notes#

Microsoft 365 token-reuse investigation

Use your identity platform’s native query language to search for session anomalies. The following generic fields illustrate useful detection logic; field names vary by provider:

SignInEvents
| where Result == "Success"
| where UserId in (suspected_users)
| summarize
    FirstSeen=min(TimeGenerated),
    LastSeen=max(TimeGenerated),
    IPs=make_set(IPAddress),
    Devices=make_set(DeviceId),
    UserAgents=make_set(UserAgent)
  by UserId, SessionId
| where array_length(IPs) > 1 or array_length(Devices) > 1

Correlate with:

  • New device or browser registrations.
  • Refresh-token issuance and reuse.
  • OAuth consent events.
  • Mailbox-rule creation or modification.
  • Conditional-access policy changes.
  • Sign-ins shortly after a phishing alert.

Phoenix remediation checks

Confirm the deployed version and identify internet exposure:

python -m pip show arize-phoenix
# or, where applicable:
docker images --format '{{.Repository}}:{{.Tag}}' | grep -i phoenix

# Check listening services on the host
ss -lntp

After upgrading, rotate relevant secrets and inspect application and reverse-proxy logs for suspicious GraphQL requests. Do not expose GraphQL or administrative interfaces directly to the public internet unless access is strongly authenticated and restricted.

CubeSandbox host-mount review

Search deployment configuration and audit data for the affected metadata key:

grep -R --line-number --fixed-strings 'host-mount' /etc /opt 2>/dev/null
grep -R --line-number --fixed-strings 'host-mount' /var/log 2>/dev/null

Review for evidence of:

  • Mounts targeting /etc, /var/lib, container-runtime sockets, application secrets, or user home directories.
  • Unexpected writes to host configuration or credential files.
  • Access to Docker, containerd, or other runtime sockets.
  • Cross-tenant process, filesystem, or network activity.

Open Object Rexx RXAPI exposure

Check the installed version and local port exposure:

rexx -v 2>/dev/null || ooRexx -v 2>/dev/null
ss -lntp | grep ':10010'
find /tmp -maxdepth 1 -type f -perm -0002 -ls 2>/dev/null

These commands are inventory checks, not proof of exploitability. Upgrade to 5.2.0, restrict access to RXAPI, and investigate unexpected processes or macros associated with REXX execution.

Incident-response evidence preservation

Before deleting suspicious sessions, tokens, mailbox rules, or host artifacts, preserve relevant evidence according to organizational policy:

- Identity-provider sign-in and audit logs
- Microsoft 365 unified audit and mailbox-rule events
- Endpoint process, browser, and network telemetry
- Reverse-proxy and GraphQL access logs
- Container-control-plane and host audit logs
- Vulnerable-package versions and deployment manifests
- Secret-rotation timestamps and affected identities

Document the time window, affected accounts and systems, containment actions, and validation results.

This article may contain affiliate links. We earn a commission on qualifying purchases at no extra cost to you.

CISA KEV additions this week#

CISA added 3 vulnerabilities to the Known Exploited Vulnerabilities catalog in the seven days to this digest. Source: CISA KEV catalog.

All KEV additions by date

Last verified: 2026-11-17

Disclaimer: This article may contain affiliate links. We earn a commission on qualifying purchases at no extra cost to you.