N035: Azure's MCP server lets agents change resources with your full rights
Azure's MCP server now lets AI agents write to cloud resources under the user's own rights, Citi shifts agent-hack losses to customers, and Claude Code moves more limits into the harness.
N035 · Friday 20 November 2026 · 5-min read · Covers 19 Nov 04:00 PT to 20 Nov 04:00 PT
Jump to: Outside agents today · Agent threats & vulnerabilities · Policy, protocols & compliance · Architecture & harnesses · Models & infrastructure · Autonomy, swarms & robotics · Alignment & safety · Defenses, tools & GitHub · Pro deep dives
Today in 30 seconds
- Microsoft's Azure Resource Manager MCP server now lets AI agents create, update and restart Azure resources with the signed-in user's own RBAC rights; an "approve" click in the client limits nothing.
- Citi now holds customers responsible for whatever a third-party AI agent does in their bank account, even after the agent is hacked; PNC and Citizens changed terms too.
- Claude Code 2.1.323 to 2.1.328 moved limits into the harness: sandboxed shell commands now ask first, and helpers run without sign-in tokens.
Outside agents today
- VMware vCenter: A suspected China-linked group exploited a vCenter authentication bypass (CVE-2026-123890, CVSS 9.8) for two months before the patch, per Mandiant (SecurityWeek, Broadcom). Upgrade to 8.0 U3m, 9.0.2.0200 or 9.1.1.0200; 7.0 gets no fix (actively exploited).
- Ruby2Shell: A heap overflow in Ruby's upper-casing lets unauthenticated attackers run code on default Rails apps using single-byte encodings (Veria Labs, GHSA-34m4-m9fh-gqp5). Upgrade to Ruby 3.3.14, 3.4.13 or 4.0.10, including GitLab and GitHub Enterprise Server hosts; exploit code is promised in a week.
Top story
Microsoft's Azure MCP server now lets AI agents change resources with the user's full rights
13 Nov 2026 · Microsoft Learn · Design risk · Public preview · No CVE
An AI assistant connected to Azure can now change and restart your cloud resources, doing anything the signed-in person may do. Microsoft's Azure Resource Manager MCP server gained create_or_update_resource and execute_resource_action (start, restart, key retrieval), each run under the user's RBAC rights. Its docs say a client approval click grants nothing and a prompt "isn't an authorization boundary", so RBAC scope is the only limit.
For leaders: Any engineer who connects an AI assistant here lends it their full Azure rights; decide whether agent sessions must use separate, narrower identities.
For builders: Run agent sessions as an identity with Reader plus Contributor on one sandbox resource group; Microsoft's sample Policy rule covers deployments only, not the new write tools.
Agent threats & vulnerabilities
AG2's restricted shell tool ran any command placed after a newline
Advisory 29 Sep 2026, CVE assigned 5 Nov (catch-up) · GitHub advisory · Vulnerability · CVE-2026-113663 · GHSA-2h4w-29cv-2cmv · CVSS 8.8 · Relevance 7
An agent framework's "safe" shell tool let a model-written command run anything, even in read-only mode. AG2's SandboxShellTool filter ignored newlines, so text after one ran as a second command under sh -c, bypassing allowed=[...] and readonly=True.
Do this: Upgrade ag2 to 1.1.1; legacy autogen 0.13.2 to 0.14.1 has no fix, so leave it or wrap the tool in an OS sandbox.
Also:
- Windows 11 logs which coding agents you run: telemetry flags claude.exe, codex.exe and cursor.exe use (one researcher's finding, 16 Nov).
Policy, protocols & compliance
Citi makes customers liable when a third-party AI agent holding their login is hacked
18 Nov 2026 · TechCrunch · Final (bank terms) · US · Relevance 8
Big US banks now say losses through an outside AI assistant are the customer's, even after a hack. Citi's updated user agreement makes customers who give third-party agents account access responsible, including for attackers using credentials an agent leaks; PNC and Citizens changed terms too.
Do this: Bar third-party AI agents from corporate bank logins, and ask your bank which agents carry no extra liability.
Architecture & harnesses
Claude Code now asks before sandboxed shell commands and strips tokens from sandbox helpers
12-19 Nov 2026 · GitHub releases · Harness release · v2.1.323 to v2.1.328 · No CVE · Relevance 7
Anthropic's coding agent now runs fewer commands without asking. Releases 2.1.323 to 2.1.328 make sandboxed Bash prompt by default, run sandbox helpers without tokens, and fix allow rules that ran a different program than named.
Do this: Update to 2.1.328 and search settings for sandbox.autoAllowBashIfSandboxed: true, which restores prompt-free runs.
Also:
- OpenAI details the Codex sandbox: Seatbelt, bubblewrap plus seccomp, and Microsoft's MXC on Windows; MCP server launches are not sandboxed by default (18 Nov).
Models & infrastructure
Google's Gemini 4 Neon is built to direct parallel coding sub-agents
19 Nov 2026 · Google · Model release · Closed, API · Relevance 7
Google's new coding model is built to manage other AI agents working on parts of a task at once. Gemini 4 Neon (closed, API) directs parallel sub-agents in Antigravity; Google reports 0.0% attack success on Gray Swan's indirect prompt-injection test (self-reported).
Do this: Scope each sub-agent's tools and check its output before the orchestrator acts.
Also:
- Microsoft's MAI Code 2: closed Copilot coding models "not yet tuned" for long agent runs, rated high on offensive cyber (self-reported).
Autonomy, swarms & robotics
Also:
- A US Army ground robot launched 18 drones that struck a target autonomously: no soldiers flew them (12 Nov demo, Telegraph).
Alignment & safety
Also:
- OpenAI lays out its frontier risk layers: no internet for research clusters; monitors pause work on a bypass (18 Nov).
Defenses, tools & GitHub
Also:
- Anthropic's tiered cyber access, measured: Defense Access blocked 46 of 50 attack trials; vetted Red Team Access completed 34 of 50 (self-reported).
One concept
Authorization boundary
An authorization boundary is where a system actually checks an action and refuses it if not allowed: an IAM role, a database permission, a sandbox. A prompt telling an agent to behave is not one, as Microsoft's Azure MCP docs say.
Basic mitigation: For every agent tool, write down which system enforces its limit and test that it refuses an out-of-scope call.
AgentSec Brief is published by AM Technology Solutions. Researched and drafted with AI agents, verified against primary sources, and reviewed before publication. Informational only; not legal, compliance or security advice. Verify before acting.
Pro members get the full deep dives with attack diagrams, detection queries and the fix in order, plus the vulnerability tracker, compliance calendar, open-source picks and watch list. Upgrade at https://www.agentsecbrief.com/#/portal/signup