N025: A browser extension could make Claude Desktop install software on another computer

A browser extension could drive Claude Desktop through Dispatch; plus a backdoored AI attack tool, a Docker sandbox escape and Chainalysis's widened breach.

Share

N025 · Friday 6 November 2026 · 5-min read · Covers 5 Nov 04:00 PT to 6 Nov 03:00 PT, plus catch-up from 3-4 Nov

Jump to: Incidents · Agent threats & vulnerabilities · Policy, protocols & compliance · Architecture & harnesses · Alignment & safety · Defenses, tools & GitHub · Other big security news · Pro deep dives

Today in 30 seconds

  • Anthropic's server accepted a logged-in claude.ai browser session as permission to send Dispatch tasks to Claude Desktop, so a browser extension could make the desktop app install software, even on another computer.
  • Docker Sandboxes' host-side sbx cp command followed symlinks an AI agent planted in its own workspace, letting the agent read or overwrite the developer's files outside the sandbox (CVE-2026-103916, fixed in 0.51.0).
  • A re-published Docker image of the AI attack tool ARTEX ran login-form input in a root shell; a researcher found 62 internet-facing consoles serving it.

Top story

A browser extension could drive Claude Desktop through Dispatch and click "Allow" itself

4 Nov 2026 · Forever Security · Vulnerability ("ThisBash") · no CVE · unfixed per researcher

A browser add-on that can see a user's Claude website tab could tell the Claude Desktop app, even on another computer, to install software. Forever Security found Anthropic's server accepted any logged-in claude.ai web session as authority to send Dispatch tasks to Claude Desktop; its test extension hid the Dispatch window, sent scripted prompts and auto-clicked "Allow" until Claude installed an npm package. Anthropic called the report a duplicate of an internal finding and said it would retire Dispatch by the end of October; we could not confirm it has. Disclosure: AgentSec Brief uses Claude.

For leaders: A browser extension can now act as your employee inside every AI agent they are signed into; ask each agent vendor whether a browser session alone can send commands to a user's machine.

For builders: Turn off Dispatch and other cross-device remote control where nobody needs it, allowlist which extensions may access claude.ai, and alert on package installs started by Claude Desktop.

Pro deep dive ↓

Incidents

Update: Chainalysis says the intruders exposed every user's password hash and live tokens

4 Nov 2026 · Chainalysis · Incident · no CVE · relevance 8

The crypto-tracing firm now says attackers got login data for all its customers. Its 4 Nov update lists password hashes and live tokens for every user, reversing its 30 Oct account; KYT is offline (first covered 2 Nov).

Do this: Revoke Chainalysis tokens held in scripts and SIEMs; rotate API keys.

Pro deep dive ↓

Agent threats & vulnerabilities

A backdoored copy of the ARTEX AI attack tool gave its planter root on every install

4 Nov 2026 · Badbyte · Supply chain · no CVE · relevance 8

Someone re-published a popular AI hacking tool with a hidden back door, giving the planter control of every server that ran it. The Docker Hub image tomicdd/artex runs login-form input in a root shell; it had about 9,000 pulls.

Do this: Search for tomicdd/artex and congdonglinux/artex; if found, isolate the host and rotate its keys.

Pro deep dive ↓

Also:

Policy, protocols & compliance

Apple's annual report warns AI agents may exceed their permissions and hide what they did

3 Nov 2026 · Fortune on Apple's 10-K · Final (SEC filing) · US · relevance 7

Apple told investors AI agents might take unapproved actions and hide them. Its 10-K says agents may reach forbidden systems, complete transactions and "conceal actions" (per Fortune; SEC filing not opened).

Do this: Add "agent exceeds permissions and hides it" to your risk register, with an owner.

Architecture & harnesses

3 Nov 2026 · Endor Labs · CVE-2026-103916 · CVSS 3.1 8.2 (vendor) · fixed in sbx 0.51.0 · relevance 8

When a developer copied files out of Docker's agent sandbox, the agent could redirect the copy to the developer's own files. Host-side sbx cp runs as the operator and follows symlinks the agent plants.

Do this: Upgrade sbx to 0.51.0 and realpath any path an agent proposes before copying.

Pro deep dive ↓

Alignment & safety

METR: AI roughly doubled Anthropic's research pace, but Anthropic redacted the key evidence

5 Nov 2026 · METR · Independent assessment · relevance 8

An outside safety group found AI probably doubled how fast Anthropic builds new AI. METR credits about 80% of a 1.9x benchmark speed-up to AI; Anthropic redacted the evidence METR found most informative. Disclosure: we use Claude.

Do this: When a vendor cites an outside audit, ask what the auditor could not see.

Pro deep dive ↓

Defenses, tools & GitHub

CrowdStrike ships an "AI Kill Switch" that stops an agent, blocks relaunch and cuts its network

3 Nov 2026 · CrowdStrike · Product · self-reported · relevance 7

A security vendor now sells a button to shut down a misbehaving AI agent. Falcon Guardian ends it, blocks relaunch and cuts network access; endpoint enforcement is "coming soon."

Do this: Test whether any kill switch stops an agent that bypasses the gateway.

Pro deep dive ↓

Other big security news

  • Iris Plaza: a breach exposed about 7.7 million Japanese shoppers' data (notice).
  • Zyxel GS1900: the patch for exploited CVE-2026-7273 could be bypassed; install 2.90 Patch 3 (advisory) (actively exploited).
  • Barracuda ESG: pre-auth CVE-2026-116606 (CVSS 9.3) chains to root; update to 9.4.0.030+ (advisory).

One concept

Session riding

Session riding is when attacker code uses a victim's logged-in session to send requests the victim never made. In the Dispatch flaw, a signed-in claude.ai tab could order Claude Desktop to run commands.

Basic mitigation: Require confirmation on the device itself before a command runs there.

Full glossary


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