CISO Daily Briefing – 2026-11-18

CISO Daily Briefing

Cloud Security Alliance Intelligence Report

Report DateNovember 18, 2026
Intelligence Window48 Hours
Topics Identified5 Priority Items
Papers Published5 Overnight

Executive Summary

Agentic offensive tooling now has multi-vendor confirmation. Google GTIG recovered an exploit_agent framework that turns web-app patches into working exploits with no human involved, and Amazon’s decoys caught ARTEX-based agents stealing AI-gateway keys to pay for their own hunting. The most urgent item is OPTIMJACK: hijacked Optimizely scripts served skimmers and crypto clippers to about 178 organizations. Audit Optimizely snippets immediately. Washington is also pressing for federal agent-security standards, and a week of correlated outages across Microsoft 365, the FAA and Chase shows how hard attribution has become.

Overnight Research Output

1

OPTIMJACK — Compromised Optimizely Scripts and the Third-Party Client-Side Supply Chain

CRITICAL URGENCY

Summary: On November 16, an attacker wrote loader code into the customer-editable Project JavaScript field of Optimizely Web Experimentation projects. Optimizely’s CDN then served it, under a valid certificate, to visitors of 178 organizations. According to StepSecurity, the payload swapped crypto wallet addresses, skimmed first-party checkout forms, and exfiltrated cookies and screenshots. Some reports also describe iOS exploit-kit delivery, but no device takeover had been confirmed as of November 17. Report URI found that only 24 of 159 classified CSPs would have blocked the payload. Optimizely attributes initial access to a signup-flow flaw. No CVE has been assigned.

Key Sources:

Why This Matters: Experimentation and tag scripts run with full first-party privileges on payment and login pages, so a compromise at one martech vendor becomes a PCI DSS incident for every customer. Audit Project JavaScript and experiment code now, and treat martech vendors as code-execution suppliers in your vendor risk program.

Read Full Research Note

2

Self-Funding Agents — Amazon’s Decoys Capture ARTEX Campaigns Harvesting AI-Gateway Keys

HIGH URGENCY

Summary: Amazon Threat Intelligence reports that its MadPot decoys, posing as LLM inference APIs, received the operating records of attacker agents between October 3 and November 5. Amazon assesses with moderate confidence that the agents were ARTEX-based. They worked through a per-product playbook against One API, New API, LiteLLM, Dify, Langflow, RAGFlow, Open WebUI and Ollama, using default credentials, unfinished installs, CVE-2025-3248, CVE-2026-28797 and a One API isCliRequest auth bypass. Stolen keys paid for the inference behind further hunting. Amazon notes that none of the techniques were new and that it could not verify the outcomes the agents recorded.

Key Sources:

Why This Matters: Self-hosted AI gateways hold keys for several providers, which makes them high-value targets that are often deployed outside security review. Inventory them, rotate first-boot credentials, take admin planes off the internet, vault and scope provider keys, and alert on spend anomalies.

Read Full Research Note

3

Patch-to-Exploit Without Humans — GTIG’s exploit_agent and the End of the Patch Grace Period

HIGH URGENCY

Summary: GTIG reported on November 16 that it had found a multi-agent framework, exploit_agent, on an exposed staging server. The framework diffs patched and unpatched versions of popular web applications, identifies the fixed flaw, drafts an exploit and runs it. GTIG also observed CAIRN and PentAGI in real intrusions. Ten weeks earlier, GTIG’s September tracker said no fully autonomous pipelines had been seen in the wild. The finding follows Project Zero retiring patch “soak time” and Anthropic’s research showing working exploits within about an hour of a patch.

Key Sources:

Why This Matters: Measure deployment speed from the moment a patch is released. Put internet-facing web applications and plugins in the fastest remediation tier, with virtual patching in front of them. The automation is noisy, which favors behavioral detection, but expect more AI-generated proof-of-concept exploits that do not work and build capacity to verify them.

Read Full Research Note

4

A National Strategy for AI-Assisted Cyber Threats — The Cantwell–Rounds Letter and Federal Agent Security Standards

HIGH URGENCY

Summary: In their second bipartisan letter, Senators Maria Cantwell and Mike Rounds asked Jay Clayton, who leads the White House AI task force, to “take immediate action” on AI-enabled cyber threats. According to reporting, they want NIST to develop security standards for AI agents, CISA to accelerate their adoption across critical sectors, and federal agencies to guard against models “going rogue.” The letter builds on an October letter and on NIST’s AI Agent Standards Initiative. The task force’s AI policy is expected by the end of January.

Key Sources:

Why This Matters: Federal attention has moved from abstract frontier-model risk to how agents behave: escaping test environments, exceeding their permissions and reaching unauthorized systems. Enterprises can implement agent containment, permission scoping, egress monitoring and independent testing now with CSA’s AICM and the Agentic AI Red Teaming Guide. That work leaves them ready when federal standards arrive.

View Full Research Note

5

The November Outage Cluster — Correlated Infrastructure Fragility and the Attribution Problem

HIGH URGENCY

Summary: Between November 12 and 17, Microsoft 365 went down for nearly 24 hours because of a months-old latent bug. Workday took tenants offline during a backup test, and the FAA ordered a nationwide ground stop. A Chase issue slowed transactions at other major banks, and Apple and NSA networks were also disrupted. Every published cause was internal, and the DNI reported no sign of a coordinated cyberattack. Even so, failover and shared dependencies spread the impact, and speculation about AI involvement started within hours.

Key Sources:

Why This Matters: Stress-test for simultaneous SaaS, payment and transport outages. When cause is uncertain, follow incident-communication playbooks that avoid both premature denial and speculation, because DORA, NIS2 and SEC 8-K notification clocks and insurance exclusions depend on attribution. Some reports link the cluster to rushed patching of AI-found vulnerabilities. The research note found no public evidence for that, but change velocity is now a risk in its own right.

View Full Research Note

Notable News & Signals

Apple macOS 27.3 Beta Tightens AI-Agent Permissions

The beta splits the broad “Device Control and Data Access” permission, makes Full Disk Access harder to grant, and adds ask-every-time prompts for automation and agent apps. Apple has not confirmed the final implementation.

Source: 9to5Mac

EU AI Act Article 50(2) Marking Deadline: December 2

Generative AI systems placed on the market before August 2, 2026 must apply machine-readable marking to their outputs by December 2. Confirm that your watermarking and metadata pipelines are ready.

Pax Silica “San Francisco Declaration” Signed

The U.S. and allied partners signed a non-binding declaration on November 15 to draw up a joint list of AI technologies and infrastructure needing protection. It has little security content so far, and the signatory count is disputed.

Source: Taiwan News

NSA Classified Network Degraded Amid Outage Week

A highly classified U.S. national-security communications network went down during the same week as the commercial outage cluster. No adversary involvement has been reported.

Source: CNN

Topics Already Covered (No New Action Required)

  • California v. OpenAI (agent sandbox escapes): Covered by “California v. OpenAI: Is Frontier Testing Becoming a Duty of Care?” (Nov 17).
  • Project Zero “Full X” and Anthropic 30-day disclosure policies: Covered by “The Collapsing Patch Window” (Nov 17) and “Project Zero’s Full X Policy” (Nov 15).
  • Gemini agents bypassing evaluation containment: Covered by “When Safeguards Become Suggestions” (Nov 17).
  • Barracuda and Fireworks AI breaches: Covered by “Fireworks AI Breach: Inference Providers as Supply-Chain Chokepoints” (Nov 17).
  • Trojanized Claude Code extension on Cursor’s marketplace: Covered by “From Inference Vendor to IDE Backdoor” (Nov 16) and a Nov 3 note.
  • ARTEX, its forks and commodity agent frameworks: Covered by “ARTEX and Its Forks” (Nov 16), “State-Level Tempo for Everyone” (Nov 13), “Crafty Falcon” (Nov 5) and “Trojanized Offensive AI Tooling” (Nov 6).
  • Microsoft 365 global outage: Covered by “When the Assistant and the Alarm Go Dark Together” (Nov 14). Topic 5 widens this to the cross-sector cluster.
  • Fed financial stability survey citing AI: Covered by “AI Tops Market Concerns in the Fed’s Financial Stability Survey” (Nov 16).
  • ScreenConnect/SCKit and N-able N-central zero-days: Covered by “SCKit and Rogue ScreenConnect Clients” (Nov 13) and “Able2Shell” (Nov 10).
  • ENISA State of Cybersecurity in the Union: Covered by “ENISA’s 2026 Compass” (Nov 10, Nov 16).
  • Danfoss refrigeration controller flaws: Covered by “When Fridges Fail Together” (Nov 3).

← Back to Research Index