CISO Daily Briefing – 2026-11-16

CISO Daily Briefing

Cloud Security Alliance Intelligence Report

Report Date
2026-11-16
Intelligence Window
48 hours
Topics Identified
5 Priority Items
Papers Published
5 Overnight

Executive Summary

The AI toolchain itself is now the attack path. A stolen Fireworks AI key led through retained coding-agent traces to a trojanized Claude Code extension on Cursor’s marketplace, and an SGLang RCE at NEAR AI showed that TEEs do not protect against flaws in the inference engine running inside them. Open-source agentic pentest frameworks led by ARTEX are now commodity attack infrastructure, hunting AI gateways for provider keys. On the policy side, ENISA says AI has industrialised vulnerability discovery, and the Fed’s survey contacts now cite AI as the top financial-stability risk (87%).

Overnight Research Output

1

From Inference Vendor to IDE Backdoor: The Fireworks–Cursor–Claude Code Extension Chain

CRITICAL URGENCY

Summary: On November 13, Cursor traced its October 29 marketplace incident to an access key it had issued to Fireworks AI, which Fireworks exposed. The attacker used the key to copy shared coding-agent traces that still held user credentials, pivoted into Cursor’s SaaS tooling, and placed trojanized Claude Code builds 2.1.312/2.1.313 in the marketplace cache. Anthropic’s own releases, the VS Code Marketplace, Open VSX and npm were not affected. Cursor warns that uninstalling the extension alone may not be enough.

Key Sources:

Why This Matters: The compromise crossed four trust boundaries: an inference subprocessor, a shared training dataset, a SaaS estate and a software distribution channel. CISOs should require redaction and retention limits on agent traces shared with vendors, inventory vendor-issued data-sharing keys, and pin IDE extensions to first-party sources.

Read Full Research Note

2

Confidential AI Under Pressure: SGLang RCE at NEAR AI and the Mior/Skyfall SEV-SNP Attacks

CRITICAL URGENCY

Summary: NEAR AI disclosed that an SGLang RCE gave an attacker code execution on an inference server, where confidential-mode requests could be observed. The attacker then stole agent chat histories and customer API keys. In the same week, the Mior and Skyfall attacks showed that a malicious host can read SEV-SNP guest memory, and that Mior can also forge attestation (AMD-SB-3059). Full mitigation depends on OEM BIOS updates that AMD targets for November–December 2026.

Key Sources:

Why This Matters: Buying “confidential inference” does not protect against flaws in the software running inside the TEE. Treat inference engines as part of the trusted computing base, require attestation verifiers to enforce minimum TCB values, and ask confidential-AI vendors how quickly they patch SGLang and vLLM.

Read Full Research Note

3

ARTEX and Its Forks in the Wild: Open-Source Agentic Pentest Frameworks as Commodity Attack Infrastructure

HIGH URGENCY

Summary: Several vendors independently reported intrusions run by forkable multi-agent offensive frameworks. Amazon MadPot captured ARTEX campaigns hunting default LiteLLM keys, funded with stolen AI keys. Cloudflare saw about 400 database-dump requests in an hour through an abused CI/CD token. Arctic Wolf tied the AntiFox fork to AWS secret theft, and Unit 42 documented Cairn breaching semiconductor and defense targets. Censys counts 1,465 exposed ARTEX consoles.

Key Sources:

Why This Matters: ARTEX is AGPL-licensed, model-agnostic and has more than 12,000 forks, so it is a class of capability rather than a single tool to signature. Detect on behavior: machine-speed request cadence, LLM-API callbacks and control-plane telemetry. Harden AI gateway credentials first, because these agents go after them before anything else.

Read Full Research Note

4

ENISA’s 2026 State of Cybersecurity in the Union: What the “Strategic Compass” Means for AI-Era NIS2 and CRA Compliance

MEDIUM URGENCY

Summary: ENISA’s biennial NIS2 Article 18 report (10 November) keeps the Union threat level at “substantial” but says frontier AI is industrialising vulnerability discovery, with the time from discovery to weaponisation approaching zero. It expects report volume to outpace triage and remediation capacity. It sets five policy directions, links NIS2, the CRA and the AI Act, and recommends that critical operators attest “clear escalation paths for AI-accelerated incidents” in annual audits.

Key Sources:

Why This Matters: CRA reporting of actively exploited vulnerabilities has been live since 11 September 2026, and the CRA applies in full from 11 December 2027. EU-exposed CISOs should size vulnerability-handling capacity for AI-speed discovery and expect 2027 supervisory audits to ask how AI-accelerated incidents are escalated.

View Full Research Note

5

AI Becomes the Fed’s Top-Cited Financial Stability Risk: Agentic Cyber Events and Systemic Contagion

HIGH URGENCY

Summary: In the November 13 Financial Stability Report, 87% of the 23 market contacts the New York Fed surveyed named AI as a salient risk, up from 50% in spring. Cyber events driven by state actors or frontier models are one of three channels they named. The Fed notes that reports of agentic systems involved in cyber events have increased since May, and it describes contagion through shared third parties, liquidity and confidence. The Fed still describes the banking system as sound and resilient.

Key Sources:

Why This Matters: A prudential regulator has now linked agentic AI threats to systemic contagion, which gives CISOs a board-ready way to frame AI security investment. Priorities: compress patch cycles, map concentrated AI and third-party dependencies, govern agents as privileged non-human identities, and rehearse incidents that hit many institutions at once.

View Full Research Note

Notable News & Signals

N-central Hotfix 2 Bypassed Within Hours

New 4EverAble2Shell/Agent2Shell chains defeat N-able’s November 10 hotfix, and 21 N-central servers were found mining Monero. Keep N-central off the internet until N-able ships a fix it states is complete.

PHP2Shell Roots Check Point Gateways, Wipes ESXi

Unpatched Mobile Access gateways were rooted and backdoored with persistence that survives restores, and at least one attacker pivoted to an ESXi host and ran a wiper. Check gateways that were patched late for implants.

UnrealIRCd AI-Found Flaw Exploited in About a Day

CVE-2026-114713, the worst UnrealIRCd flaw in 15 years and found with AI help, was being exploited about a day after the hot-patch shipped. CISA added it to KEV on November 13.

Source: oss-security

PostgreSQL Patches 52 Flaws, Including CVE-2026-6460

A type confusion dating to 2004 lets users who can create objects run code as the database’s OS user. Patch, rebuild C extensions and confirm fix dates with managed-database providers.

Topics Already Covered (No New Action Required)

  • PHP2Shell exploitation (Check Point gateways, ESXi wiper): Covered in Frontier Ready daily briefings (Oct 30, Nov 14). The AI-discovery angle was covered at disclosure.
  • ARTEX / Storm-0940 initial reporting (Wiz, Oct 22): Covered in Frontier Ready briefings (Nov 4, Nov 12). Topic 3 adds the new multi-vendor evidence of in-the-wild use.
  • LiteLLM gateway compromise and exploitation: Covered by existing CSA research notes on the LiteLLM vulnerability chain and on active exploitation via MCP injection.
  • CAISI/UK AISI assessments of GLM-5.3 and Kimi K3: Covered in late-October Frontier Ready briefings. The sources are now more than eight weeks old.
  • CRA Single Reporting Platform launch: Covered in earlier briefings and ENISA CRA readiness notes (launched Sep 11).
  • Windows SignDrop zero-day, N-able N-central, Red Hat Quay, PostgreSQL CVE-2026-6460, UnrealIRCd KEV: High severity but not AI-specific. Tracked in the daily briefing rather than as standalone research notes.

← Back to Research Index