CISO Daily Briefing
Cloud Security Alliance Intelligence Report
Executive Summary
A quiet weekend for new AI security material, but five gaps matter. The AI supply chain is the headline: MCP server dependency drift turns every coding-assistant workstation into an unreviewed install point, and TypoPrompting uses 186 fake AI-brand domains to deliver malware and leak employee searches. Project Zero’s Full X policy cuts public disclosure to 60 days in February 2027 and 30 days in 2028, so vendor SLAs and non-patch mitigations need rework now. Agentic auto-remediation and agent payment wallets are scaling faster than accountability for their failures.
Overnight Research Output
Pinned but Not Frozen: Dependency Drift in MCP Servers Launched by AI Coding Assistants
HIGH URGENCY
Summary: Most local MCP servers are launched, not installed: a config line such as npx -y server@1.4.2 pins only the top-level package, while its dependency tree resolves fresh against semver ranges. Corelight measured one commit-pinned server running with 15 distinct dependency sets across 19 builds in five weeks. The March 2026 axios compromise shows how one poisoned transitive release reaches every resolution made in its window. Pin exactly, add release-age cooldowns, disable install scripts, and move high-value servers to digest-referenced images.
Key Sources:
TypoPrompting: Lookalike AI-Brand Domains Delivering ClickFix, Hijackers and Search-Leaking Extensions
HIGH URGENCY
Summary: Island documented 186 typo domains imitating Claude, ChatGPT, DeepSeek, Perplexity and others. Fingerprinting gates send Windows users to fake Cloudflare checks that run hidden PowerShell (ClickFix); other paths deliver Net Guard, byte-identical to a Microsoft Store listing, or Chrome extensions that override the default search provider while requesting no permissions. The domain count is an inventory, not a confirmed victim count. Publish managed launchers for sanctioned AI tools, block indicators, enforce default search and extension allowlists, and hunt for browser-to-Run-dialog execution.
Key Sources:
Project Zero’s “Full X” Disclosure Policy: Governance Implications of a 30-Day World
HIGH URGENCY
Summary: Google Project Zero replaced its 90+30 model, arguing that with LLMs “zero patch-to-exploitation time is a reasonable, though imperfect, model of attacker behavior.” Grace periods, variant rules and severity-based timing are gone. Enterprises, not vendors, will feel this first: details of flaws in widely deployed products may be public about 30 days after vendor notification, patched or not. Ask suppliers for evidence of feature-flag, filtering and hotpatch capability, and reconcile patch-gap targets with EU CRA reporting duties.
Key Sources:
Who Owns the Bad Patch? Security Risks of Agentic Auto-Remediation at Scale
MEDIUM URGENCY
Summary: GitHub agentic autofix reached GA on 15 October, and Snyk and OpenAI Daybreak are expanding similar agents. Evidence lags adoption: one study found only 24.8% of 319 LLM security patches fully correct (0% for input-validation flaws), and another confirmed 170 silent failures that tests and LLM reviewers missed. Assign ownership of AI-authored fixes, cap PR volume to review capacity, and scope agent write access before scaling.
Key Sources:
Agents at the Checkout: Loss Allocation and Gatekeeper Concentration in Frontier-Lab Payment Wallets
MEDIUM URGENCY
Summary: OpenAI launched ChatGPT Wallet on 13 November with American Express, Mastercard and Visa. Neither the announcement nor merchant documentation says who pays when a manipulated or mistaken agent completes an approved purchase. Because Regulation E turns on whether a transaction was “unauthorized,” biometric approval likely shifts error costs away from issuer dispute processes. The lab now sits at several control points: agent, signing keys, consent screen and conversation record.
Key Sources:
Notable News & Signals
SignDrop: Public Exploit for Unpatched Windows LPE
A public exploit chains Windows CLFS and Defender for local privilege escalation; no in-the-wild use reported yet. Watch for exploitation and an out-of-band fix.
Google: LLMjacking Bills Reach Up to $1 Million
Google warns that stolen cloud credentials used to run AI models can generate charges up to $1M per victim; set billing alerts and quota caps on AI services.
OpenAI Expands Daybreak Agentic Vulnerability Fixing
OpenAI broadened Daybreak’s agentic vulnerability-fix offerings, adding to the remediation-agent wave covered in Topic 4.
EU CRA Vulnerability Reporting Burden Comes Into Focus
Practitioner commentary on CRA actively-exploited-vulnerability reporting, which now interacts with compressed public-disclosure timelines.
Vulnerability Backlogs Outpace Remediation Capacity
New data shows security teams’ backlogs growing faster than fixes ship, sharpening the case for non-patch mitigations.
Topics Already Covered (No New Action Required)
- Microsoft 365 outage (Copilot + Defender down together): CSA research note on assistant and alarm going dark together (14 Nov).
- Stolen Claude API keys as coding-agent implants: Covered by 13 Nov and 8 Nov research notes, including the Claude Code settings secret leak.
- Post-training backdoors in coding agents (Microsoft research): Covered 14 Nov and 11 Nov.
- Unit 42 Cairn autonomous AI intrusions: Covered by the agentic capability diffusion to state actors note (13 Nov).
- MCP Inspector OAuth RCE (CVE-2026-119025): In the CSA research index and Frontier Ready 7 Nov briefing.
- AMD SEV-SNP Mior/Skyfall attestation breaks: Covered 14 Nov.
- PHP2Shell on Check Point gateways; N-central 4EverAble2Shell: Covered by existing exploitation notes (10 Nov and prior).
- Apache mod_http2 RCE; GitHub CI secret stealer: Covered 12 Nov.
- Deepfake fake-Zoom intrusions (North Korea): Covered 14 Nov.
- AI agent kill-switch incident response; IBM sovereignty audit-trail report: Incremental to existing kill-switch, incident-response and sovereignty coverage.