CISO Daily Briefing – November 15, 2026

CISO Daily Briefing

Cloud Security Alliance Intelligence Report

Report Date
November 15, 2026
Intelligence Window
48 Hours
Topics Identified
5 Priority Items
Papers Published
5 Overnight

Executive Summary

Today’s five research notes cover risks that sit around AI models, not inside them. MCP servers launched via npx/uvx pull fresh, unreviewed dependencies on every start, and TypoPrompting uses 186 lookalike domains for 27 AI brands to deliver ClickFix lures and search-leaking extensions. Google Project Zero’s Full X disclosure policy drops to a 60-day deadline on 2 February 2027 and to 30 days in January 2028, which will compress enterprise patch windows. AI auto-remediation and ChatGPT Wallet both create new paths for code and money that current accountability controls do not cover.

Overnight Research Output

1

Pinned but Not Frozen — Dependency Drift in MCP Servers Launched by AI Coding Assistants

HIGH URGENCY

Summary: Corelight measured that 23 of 35 MCP servers resolve dependencies from public registries every time they launch. One commit-pinned server ran with 15 different dependency sets across 19 builds in five weeks. npm never publishes lockfiles, and an exact uvx pin locks only the top-level tool. Every developer workstation therefore becomes a recurring, unreviewed install point that runs registry code with the developer’s credentials. The March 2026 axios compromise showed that a malicious transitive release can reach every fresh resolution within hours. CSA recommends exact pins, release-age cooldowns, disabled install scripts, and digest-referenced container builds for high-value servers.

Key Sources:

Why This Matters: Configuration-level MCP allowlisting cannot detect drift because the code changes without any edit to the config file. Pinning also freezes vulnerable versions (as with CVE-2025-6514), so pins need a tracked update process with a named owner.

Read Full Research Note

2

TypoPrompting — Lookalike AI-Brand Domains Delivering ClickFix, Browser Hijackers and Search-Leaking Extensions

HIGH URGENCY

Summary: Island researchers catalogued 186 lookalike domains that route visitors through IP fingerprinting into three paths. The first is a fake Cloudflare check that prompts users to paste a hidden-PowerShell command. The second is Net Guard, a browser hijacker byte-identical to a Microsoft Store listing. The third is a family of Chrome extensions that override the default search provider. Those extensions request no permissions and run no JavaScript, yet they send every address-bar search to a third party, which can expose internal hostnames, customer names and project details. Island counted domains, not victims, and has published no infection or loss figures.

Key Sources:

Why This Matters: Unmanaged access to AI tools is now a data-leakage channel as well as an initial-access vector. Store signatures and permission prompts did not catch these cases, so controls must inspect search-provider overrides and look for browser activity followed by Run-dialog or Terminal execution.

Read Full Research Note

3

Project Zero’s “Full X” Disclosure Policy — Governance Implications of a 30-Day World

HIGH URGENCY

Summary: On 13 November Google Project Zero replaced its 90+30 policy with fixed full-disclosure deadlines. The stated reason is that LLMs have collapsed patch-to-exploit time, so patch soak time “is no longer meaningful.” Grace periods, variant rules and severity-based timing are all removed. Vendor severity ratings will therefore matter less for prioritization, because low-rated bugs become public on the same schedule as critical ones. Enterprises lose the post-patch adoption buffer and should ask suppliers for feature-flag, filtering and hotpatch capabilities. The policy also interacts with EU CRA reporting duties and CISA KEV remediation deadlines.

Key Sources:

Why This Matters: Project Zero sets norms for coordinated disclosure, and other research teams are already shortening their timelines. TPRM, procurement and patch governance should plan for a 30-day norm, not treat it as a Google-specific exception.

View Full Research Note

4

Who Owns the Bad Patch? Security Risks of Agentic Auto-Remediation at Scale

MEDIUM URGENCY

Summary: GitHub Agentic Autofix reached general availability and lets teams assign batches of alerts to the Copilot coding agent. Snyk, Legit Security, Wiz and OpenAI Daybreak are moving the same way. The evidence on correctness is weak. One study found only 24.8% of LLM-generated security patches fully correct, with 0% success on input-validation flaws. Another confirmed 170 silent failures that neither tests nor LLM reviewers caught. Remediation agents also hold write access across many repositories while reading untrusted code and issues, which makes them a new privileged identity that prompt injection can target.

Key Sources:

Why This Matters: Vendors say customers decide what merges, so accountability lands on the deploying organization. Set review-gate capacity, provenance tagging, “fix introduces new CWE” regression tests and model-change control before running fleet-wide campaigns. Outcome metrics should replace MTTR.

Read Full Research Note

5

Agents at the Checkout — Loss Allocation and Gatekeeper Concentration in Frontier-Lab Payment Wallets

MEDIUM URGENCY

Summary: ChatGPT Wallet lets OpenAI’s agent complete merchant purchases with network-issued tokens after a biometric check, so neither the agent nor the merchant sees the card number. OpenAI documentation does not say who pays when an agent buys the wrong item or is prompt-injected into a purchase. Under Regulation E and Regulation Z, a per-purchase biometric approval is strong evidence that the cardholder authorized the transaction. The frontier lab also holds the signing keys merchants use to admit agents and the conversation record that decides disputes. CSA expects agent-error losses to flow to consumers and merchant returns rather than issuer fraud processes.

Key Sources:

Why This Matters: Enterprise procurement and travel agents will get similar rails soon. Treat agent-initiated purchases as a distinct transaction class in fraud, dispute and expense controls. Merchants should decide deliberately which agent signatures to admit rather than inherit a CDN default.

View Full Research Note

Notable News & Signals

SignDrop: Unpatched Windows CLFS Privilege Escalation Exploit Public

A public exploit forges CLFS signature files to take a standard user to SYSTEM via Defender. It was not fixed in November’s Patch Tuesday, and no in-the-wild use has been reported yet.

Google: LLMjacking Bills Reach Up to $1 Million per Victim

Google’s GTIG and Mandiant say stolen cloud identities drive LLMjacking, with some victims charged over $100,000 a day. Set spend caps and revoke exposed keys quickly.

OpenAI Expands Daybreak as the Cyber-Defense Window Narrows

OpenAI broadened Daybreak’s agentic vulnerability-fix offerings. This is another sign that AI-written patches are heading into production pipelines at scale.

Source: OpenAI

EU CRA Vulnerability Reporting: Practical Burden Comes Into Focus

Practitioner commentary on the Cyber Resilience Act’s 24-hour actively-exploited-vulnerability reporting duty, which now overlaps with shrinking disclosure timelines.

Vulnerability Backlogs Keep Growing Faster Than Teams Can Fix

Industry data shows security teams falling further behind on remediation backlogs, just as AI-driven discovery and 30-day disclosure norms push volume up.

Topics Already Covered (No New Action Required)

  • Microsoft 365 outage (Copilot + Defender, 12 Nov): Covered by When the Assistant and the Alarm Go Dark Together (14 Nov).
  • Stolen Claude API keys as coding-agent implants: Covered by the 13 Nov coding-agent API key implant note; the related Claude Code settings secret leak is covered in the 8 Nov note.
  • Microsoft research on hidden backdoors in AI agents: Covered by the post-training backdoors in coding agents (14 Nov) and Lazarus neuron-edit backdoor (11 Nov) notes.
  • Unit 42 Cairn autonomous AI intrusions: Covered by the 13 Nov note on agentic capability diffusion to state actors.
  • MCP Inspector / open package OAuth RCE (CVE-2026-119025): Already in the CSA research index and the Frontier Ready 7 Nov briefing.
  • AMD SEV-SNP Mior/Skyfall: Covered by the 14 Nov confidential-AI attestation note.
  • PHP2Shell on Check Point gateways (ESXi wiper follow-on): Covered by the PHP2Shell pre-disclosure exploitation note.
  • N-able N-central 4EverAble2Shell/Agent2Shell: Covered by the 10 Nov N-central zero-days note.
  • Apache HTTP Server mod_http2 RCE: Covered by the 12 Nov mod_http2 note.
  • GitHub CI secret stealer: Covered by the 12 Nov WFX GitHub CI secret stealer note.
  • North Korean deepfake fake-Zoom scheme: Covered by the 14 Nov FaceStealer deepfake meeting intrusion note.
  • AI agent kill-switch incident response: Covered by the AI kill-switches governance paper and the 27 Oct agentic incident response note.
  • Digital sovereignty audit-trail findings (IBM): Adds little to recent CSA sovereignty coverage (OSERA, EU tech sovereignty, AI vendor sovereign dependency).

← Back to Research Index