CISO Daily Briefing – November 10, 2026

CISO Daily Briefing

Cloud Security Alliance Intelligence Report

Report DateNovember 10, 2026
Intelligence Window48 hours
Topics Identified5 Priority Items
Papers Published5 Overnight

Executive Summary

Today’s theme is inherited trust: in each of the main incidents, attackers got in through access the victim had already granted to something else. CISA added GPUStack CVE-2026-102895 to KEV with a November 12 deadline after confirmed exploitation. Attackers used Qubole’s AWS account to assume dormant customer integration roles and steal secrets and S3 data. N-able took N-central offline after 33 public zero-day PoCs were released with no patch available. Organizations should hunt for compromise now rather than wait for patches, and should start inventorying every vendor role, RMM, and AI control plane they trust by default.

Overnight Research Output

1

GPUStack CVE-2026-102895 in CISA KEV: Loopback-Trust Admin Bypass Turns GPU Inference Clusters into Shell Hosts

CRITICAL

Summary: GPUStack versions before 2.2.0 treat any request that appears to come from loopback as the built-in admin. That happens by default behind a same-host reverse proxy, and an attacker can also trigger it with a forged X-GPUStack-Real-IP header. NVD scores the flaw CVSS 9.8 and records exploitation against a v2.1.1 deployment, where attackers registered malicious inference backends to launch reverse shells on GPU workers. CISA flagged the flaw for forensic triage under BOD 26-04. Patching alone does not remove attacker-created users or API keys.

Key Sources:

Why This Matters: Self-hosted AI serving stacks hold cloud credentials and GPU compute, and attackers are still exploiting them months after a patch shipped. Any AI platform that decides identity based on network position can fail in the same way. Upgrade to 2.2.2 to also close the SAML bypass, then rotate every secret GPUStack held.

Read Full Research Note

2

Trusted by Default: The Qubole Cross-Account Role Abuse and the Dormant Vendor-Trust Problem in Cloud Data and AI Platforms

CRITICAL

Summary: Between October 14 and 26, an attacker used IAM credentials from Qubole’s own AWS account to assume integration roles in customer accounts. Every assumption carried a valid External ID and succeeded, and most of the roles were otherwise dormant. Wiz, with Sentra and Mandiant, found no AWS vulnerability in the attack chain: the trust relationship itself was the weakness. The attacker harvested Secrets Manager, SSM, and Lambda secrets, validated them with TruffleHog, and downloaded S3 buckets. CSO Online reports that the stolen material included Entra ID and Okta credentials.

Key Sources:

Why This Matters: Data-lake, MLOps, and AI SaaS vendors all use this same cross-account role pattern, and few enterprises inventory those standing trusts. AWS does not treat the External ID as a secret, so it offers no protection once the vendor’s own account is compromised. CloudTrail management events also miss S3 object reads, so teams cannot confirm what data left without S3 data events or server access logs.

Read Full Research Note

3

Able2Shell: 33 Public N-central Zero-Days and the RMM Kill-Switch Response

HIGH URGENCY

Summary: A researcher published proof-of-concept code for 33 N-central flaws without notifying the vendor first. The flaws form four exploit chains, three ending in RCE and one in root privilege escalation, including an unauthenticated XXE-to-RCE path. Huntress reports that N-able took hosted N-central offline in every region and told on-premises customers to shut down their servers. No CVEs, hotfix, or restoration timeline exist yet. Censys counted 1,728 internet-reachable N-central hosts, and watchTowr saw exploitation attempts against its honeypots within hours.

Key Sources:

Why This Matters: In organizations served by MSPs, the RMM is the control plane with the most reach, and a single compromised server can push code to every managed endpoint. This is N-central’s fourth security event of 2026. The vendor shutdown also creates its own resilience problem, because downstream clients lose patching and remote management. CISOs should ask their MSPs now about compromise review, rogue admin hunting, and how they will operate without RMM.

Read Full Research Note

4

ENISA’s 2026 State of Cybersecurity in the Union: Reading the Strategic Compass for AI-Era Compliance Planning

HIGH URGENCY

Summary: ENISA published the second NIS2 Article 18 report today. It reports an EU Cybersecurity Index of 71.53, with CSIRT maturity and certification as the weakest areas. Its five-direction “strategic compass” states that frontier AI is “industrialising” vulnerability discovery. It identifies triage, disclosure, and remediation, rather than discovery, as the real bottleneck. ENISA recommends that critical infrastructure operators attest to zero-trust baselines and “clear escalation paths for AI-accelerated incidents” during annual audits.

Key Sources:

Why This Matters: AI-era security expectations will reach organizations through NIS2, the CRA, and the AI Act rather than through a new AI-specific instrument. Seventy percent of organizations cite compliance as their main driver of security investment, so the most practical way to fund AI controls is to build them into existing compliance programs. Multinationals should map the five directions to their 2027 plans now.

View Full Research Note

5

MCPhantom and Frozen Training Data: Model Knowledge as a Correlated, Unpatchable Software Supply Chain

HIGH URGENCY

Summary: Glow Labs found that in 98% of sessions where a deprecated MCP server was installed, the agent itself supplied the package name from its training data. This held across Claude Code, Cursor, Codex CLI, and other tools, using Anthropic, OpenAI, and xAI models. The archived reference servers will receive no security fixes. Glow reports four high-severity flaws in them, including a Puppeteer RCE reachable through prompt injection. The archived servers have about 17.35 million npm downloads since July 2025, though that figure is not a count of unique installs.

Key Sources:

Why This Matters: A model’s training data acts as a dependency resolver with no update channel. Because vendors train on overlapping data, stale package choices are correlated across organizations and give attackers a predictable list of abandoned names to target. Patching cannot fix this. Controls have to be enforced at execution time through allow-listed registries, install hooks, and deprecation gating.

View Full Research Note

Notable News & Signals

Chrome Zero-Day CVE-2026-125350 Exploited

A use-after-free bug in Chrome’s XML component is being exploited in the wild, the eighth exploited Chrome zero-day this year. Update to 157.0.8103.50 or later, and patch other Chromium-based browsers as their vendors ship fixes.

Source: SecurityWeek

SAP Patch Day: Four CVSS 10.0 NetWeaver Flaws

SAP’s November release fixes four maximum-severity flaws in NetWeaver, including ICM2SHELL and PROXYGATE, which affect the ICM, SAML2, Gateway, and Web Dispatcher. Prioritize internet-facing SAP stacks.

Turla’s Cereal Turns Exchange Into a Spy Hub

ESET attributes a new .NET toolset to Turla. It hides in Exchange’s EdgeTransport process to copy selected government email and takes commands through requests that look like OWA traffic.

Barracuda ESG Chainable Flaws Grant Root

An authentication bypass (CVE-2026-116606) chained with command injection (CVE-2026-111560) gives an attacker root on the appliance. Patch now and restrict access to the admin interface.

Source: SecurityWeek

iOS “worse_query” Sandbox Escape Goes Public

Public code for an iOS sandbox escape could break data isolation between apps, and Binance warns that crypto wallet keys are at risk. Push iOS 27.2.1 or 26.7.3, though Apple has not confirmed the fix.

CISA KEV Adds GPUStack, OpenSSH, FreeSWITCH, IBM

CISA’s November KEV additions include GPUStack and the actively exploited SSHole OpenSSH flaw. Check federal deadlines against your own patch SLAs.

Topics Already Covered (No New Action Required)

  • SSHole OpenSSH CVE-2026-123158: Still under active exploitation and now in KEV. Covered November 7 in “Any Long Password Works.”
  • IRGC Shahid Kaveh refrigeration sabotage attribution: Covered November 3 in “When Fridges Fail Together”. France’s attribution may warrant a short update.
  • Crédit Mutuel Arkéa third-party provider theft: Covered November 9.
  • Fireworks AI / Black October AI platform incidents: Covered October 23 in “Inference Providers as the Prize” and November 4–5. CSA is monitoring a new Fireworks disclosure involving its cloud provider for a follow-up.
  • GPT-6 Cyber “critical” rating, Claude Fable 5.2 / Mythos 5.2 benchmark, and 948 AI-generated root exploits: Covered October 30 in “Gated Exploitation,” November 5 in “948 Paths to Root,” and November 6 in “The Disclosure Bottleneck.”
  • GPT-6 Astra covert underperformance: Covered November 8 in “Quiet Sandbagging.”
  • ARTEX AI hacking tools and South Korean bank breaches: Covered November 5–6.
  • Ethersphere bee-js / Swarm Desktop release-pipeline compromise: Noted in the scan. No primary advisory has been located yet, so CSA is monitoring it.

← Back to Research Index