Detection queries Qualified report

Forever Security EMBRACE, browsers embedded in five AI assistants

Forever Security (Gal Weizman; published Oct 29, 2026) calls the shared pattern EMBRACE, an embedded browser trusted like a standalone one. The post covers Cursor, GitHub Copilot, Codex, VS Code, and Claude…

What happened

Forever Security (Gal Weizman; published Oct 29, 2026) calls the shared pattern EMBRACE, an embedded browser trusted like a standalone one. The post covers Cursor, GitHub Copilot, Codex, VS Code, and Claude Desktop. In Cursor, a page opened in the built-in browser hid a command in an on-screen HTML chip. The agent read the hidden instructions. In an otherwise empty composer, one Enter ran the proposed command. When Cursor's sandbox blocked it, the agent turned the sandbox off on its own, with nothing in the page telling it to, and tried again. Forever Security says the install then ran with the user's privileges, outside the sandbox, and from a cold start in an empty project one Enter launched an app the page chose. The post says this does not fire on every model. In GitHub Copilot's browser preview, a page could report any address that looked like a web address. The address bar and the agent's tab list both showed github.com while the tab was somewhere else. Forever Security confirmed that on Copilot 1.1.27 on Sep 26, 2026. GitHub fixed it in 1.1.29 on Oct 13, before a report was sent. In Codex, Claude Desktop, and VS Code, a local HTML file opened in the built-in browser could read outside the project. The delivery in the post is a dull financial-report repo. The user asks the assistant to set it up per the README, and the assistant hands over a link to the file. Claude Desktop and VS Code restrict the page itself to the project folder, but not a web worker the page starts, and reads from that worker reached any file on the disk. Codex had no folder check. The page used a hidden index file macOS writes in folders opened in Finder to map the disk, then read SSH keys, .env files, npm tokens, and the hosts file. Disclosure: Cursor, Sep 10, patched silently on Oct 1, with no reply. Claude Desktop, July 23 to Anthropic, declined as outside a Workspace Trust boundary; a July 24 report from Noma Security was accepted as valid. Codex, July 27 to OpenAI, no reply, and still unpatched at publication. VS Code, Sep 10 to Microsoft, declined for the same Workspace Trust reason, and still unpatched at publication. This is a different Forever Security report from forever-agents-md-hijack-2026 and thisbash-claude-dispatch-rce-2026-11.

Cursor, GitHub Copilot, Codex, VS Code, and Claude Desktop each ship a browser inside the assistant. Forever Security says the enrichments that let the agent use that browser broke three ordinary browser rules.

Read the original source

How it happened

Select a stage to explore its defenses
Detect this stage
Detect this stage
Select a stage · Drag to explore · + / − to zoom · Dashed edges are loopsSelect a stage · Arrows show reported connectionsResearch / qualified account
Read the flow as text & sources

A catalog reconstruction of the supplied source account. The post is a disclosure of research findings in product browsers. It does not name an in-the-wild intrusion. At publication, Forever Security said the Codex and VS Code issues were still unpatched.

  1. 01 · Page or local file opens in the assistant browser

    A site is opened in Cursor or Copilot, or the assistant links a local HTML file from a repository the user asked it to set up.

    Hunt: Repeated exploit or upload requests are followed by an unexpected server process, file, or outbound connection.

    A scan or high request rate alone does not prove compromise or AI use. Host telemetry is needed to confirm execution.

    Source for step 1
  2. 02 · The embedded browser breaks a browser rule

    Cursor lets the page feed the agent. Copilot accepts an address the page supplies. Codex, Claude Desktop, and VS Code let a local file read outside its folder, through a web worker in the last two.

    Hunt: An agent ingests external content or an untrusted session identifier before a tool action or configuration change outside the approved task.

    Requires application instrumentation and a custom field mapping. Input exposure does not prove that the model followed instructions, executed a tool or bypassed approval. Avoid collecting sensitive prompt bodies unless authorized.

    Source for step 2
  3. 03 · Command, file read, or spoofed address

    Cursor's agent can disable its own sandbox after a block. The local file reads SSH keys, .env files, npm tokens, and the hosts file. Copilot's bar can show github.com for a different page.

    Source for step 3
  4. 04 · Two issues still open at publication

    Cursor was patched on Oct 1 and Copilot in 1.1.29. Forever Security says Codex and VS Code were still unpatched on Oct 29, 2026.

    Source for step 4

Why it matters

Reported entry point: A page opened in the assistant's built-in browser, or a local HTML file the assistant links from a repository the user asked it to set up. Reported impact: Cursor was patched on Oct 1 and Copilot in 1.1.29. Forever Security says Codex and VS Code were still unpatched on Oct 29, 2026; reported targets include Users of those five desktop assistants. No victim organization is named. The file-read demonstration used the researchers' own machine; applicability to your setup has not been assessed, so verify affected versions, permissions, and configuration in the original report.

Read the original report

What you can do

Protection & prevention

Treat the browser inside each assistant as an endpoint that loads untrusted pages and local files. Keep Cursor and Copilot on builds that include the Oct 1 and 1.1.29 fixes. Do not open local HTML from a repository you do not trust, including a file the assistant links. Do not treat a required Enter or a single click as a control when the assistant is the one asking for it.

Catalog defensive guidance based on the reported behavior. Confirm scope and current mitigations in the original source.

Review the original guidance
Investigate each phase

Detection queries

Draft query starters grouped by the attack stages where they apply.

Download stage pack

Start with the logs you collect. Each hunt identifies its required visibility, field mapping and limits. Queries are catalog-authored drafts using reusable behavior profiles; test with known activity and your normal baseline before alerting.

01Page or local file opens in the assistant browserWeb exploitation → host activityCustom schema · Draft

A site is opened in Cursor or Copilot, or the assistant links a local HTML file from a repository the user asked it to set up.

What to look for

Repeated exploit or upload requests are followed by an unexpected server process, file, or outbound connection.

Visibility: Defender-side logs where collected

Required logs

  • Reverse-proxy / WAF request logs
  • Server process and file events
  • Outbound DNS / proxy records

Coverage limit. A scan or high request rate alone does not prove compromise or AI use. Host telemetry is needed to confirm execution.

Custom-schema candidates, not out-of-box detectors. Map your application / cloud / platform logs to EventType, Actor, Asset, Action, Target and CorrelationId. KQL uses a custom OpenAttacksAudit_CL table; SPL uses openattacks:audit. EventType values are proposed normalization names, not native vendor event names. Correlation and environment-specific alert logic remain to be built.

Sigma
Draft SigmaDownload Sigma
title: Hunt - Web exploitation → host activity
id: 04d900b5-0c04-46f2-a13a-91f5574fae98
status: experimental
description: Repeated exploit or upload requests are followed by an unexpected
  server process, file, or outbound connection.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
  - Authorized activity matching this broad hunt; baseline and investigate
    before alerting.
level: low
logsource:
  category: application
  product: openattacks_audit
  definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
    Target and CorrelationId from the required source logs first.
detection:
  selection:
    EventType:
      - web.exploit_candidate
      - web.upload
  condition: selection
KQL
Draft KQLDownload KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("web.exploit_candidate", "web.upload")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated desc
SPL
Draft SPLDownload SPL
sourcetype="openattacks:audit"
EventType IN ("web.exploit_candidate", "web.upload")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _time
Investigation steps
  1. Join requests to the affected host and process tree by time and request ID.
  2. Inspect uploaded files and server-spawned shells; distinguish authorized testing from exploitation.
02The embedded browser breaks a browser ruleUntrusted content enters an agent sessionCustom schema · Draft

Cursor lets the page feed the agent. Copilot accepts an address the page supplies. Codex, Claude Desktop, and VS Code let a local file read outside its folder, through a web worker in the last two.

What to look for

An agent ingests external content or an untrusted session identifier before a tool action or configuration change outside the approved task.

Visibility: Application / agent audit records

Required logs

  • Input provenance, source type and request/session identifier
  • Captured tool responses or retrieved content where collection is approved
  • Subsequent tool calls, configuration changes and approval decisions

Coverage limit. Requires application instrumentation and a custom field mapping. Input exposure does not prove that the model followed instructions, executed a tool or bypassed approval. Avoid collecting sensitive prompt bodies unless authorized.

Custom-schema candidate for input exposure only. Map agent.input.ingest and agent.session.resume to your application request and session logs. CorrelationId must link subsequent actions. EventType values are proposed normalization names, not native vendor events.

Sigma
Draft SigmaDownload Sigma
title: Hunt - Untrusted content enters an agent session
id: 2d89d7ae-9d44-4dc6-a95b-f9c8de7b8352
status: experimental
description: An agent ingests external content or an untrusted session
  identifier before a tool action or configuration change outside the approved
  task.
author: MagicSword OpenAttacks catalog
date: 2026-11-04
falsepositives:
  - Authorized activity matching this broad hunt; baseline and investigate
    before alerting.
level: low
logsource:
  category: application
  product: openattacks_audit
  definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
    Target and CorrelationId from the required source logs first.
detection:
  selection:
    EventType:
      - agent.input.ingest
      - agent.session.resume
  condition: selection
KQL
Draft KQLDownload KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("agent.input.ingest", "agent.session.resume")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated desc
SPL
Draft SPLDownload SPL
sourcetype="openattacks:audit"
EventType IN ("agent.input.ingest", "agent.session.resume")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _time
Investigation steps
  1. Join the ingested input to later actions by session or request ID; preserve the original content as evidence.
  2. Compare the resulting action with the approved task and baseline. Receiving suspicious text alone is not successful injection.
Technical detailsTooling, impact & MITRE ATLAS

Attack context

AI's role
Agent / MCP as target
Initial access
A page opened in the assistant's built-in browser, or a local HTML file the assistant links from a repository the user asked it to set up

Tooling as reported

AI products, models and artifacts

Targets and reported impact

Targets / sectors
Users of those five desktop assistants. No victim organization is named. The file-read demonstration used the researchers' own machine.
Reported scale / exposure
Five products. Two issues, Codex and VS Code, were still unpatched when the post was published on Oct 29, 2026. Cursor was patched on Oct 1. Copilot was fixed in 1.1.29.

No target country is named. The macOS index file is a technique detail, not a victim region.

Reported attribution

No threat actor. Forever Security disclosed the findings to the vendors. Noma Security is credited for a July 24 Claude Desktop report that Anthropic accepted.

Structured attribution

Actor
unknown
Motivation
unknown
Attributed by
Forever Security
Attribution confidence
low

Detection rules

  • Embedded-browser findings have no single host ruleGUIDANCE

    The post does not name a process, a package, or a network destination that a Sigma, Splunk, or Elastic rule would match. Hunt assistant browsers opening local HTML, an agent disabling its own sandbox after a blocked command, and Copilot preview tabs whose recorded address is github.com while the page is not. The macOS folder index is a file the page reads, not a separate malware family.

Activity timeline

First seen
2026-07-23
Last updated
2026-10-29

Updates

  • 2026-07-23 — Forever Security

    Claude Desktop report to Anthropic, declined as outside Workspace Trust. Noma Security's July 24 report was accepted as valid. Codex was reported to OpenAI on July 27.

  • 2026-09-10 — Forever Security

    Cursor and VS Code reports. Microsoft declined the VS Code issue for Workspace Trust.

  • 2026-10-01 — Forever Security

    Cursor silently patches its issue. Forever Security says Cursor did not reply.

  • 2026-10-13 — Forever Security

    GitHub fixes the Copilot address-bar spoof in 1.1.29, before a report was sent.

  • 2026-10-29 — Forever Security

    Gal Weizman publishes EMBRACE. Codex and VS Code are described as still unpatched.

Source defensive notes

Treat the browser inside each assistant as an endpoint that loads untrusted pages and local files. Keep Cursor and Copilot on builds that include the Oct 1 and 1.1.29 fixes. Do not open local HTML from a repository you do not trust, including a file the assistant links. Do not treat a required Enter or a single click as a control when the assistant is the one asking for it.

MITRE ATLAS

AI-specific catalog annotations based on the reported behavior; these are not MITRE case-study assignments.

AML.T0051.001Indirect

A page or local file the assistant opens carries instructions or a command the agent then acts on Source

AML.T0055Unsecured Credentials

The local-file finding read SSH keys, .env files, and npm tokens from the researchers' machine Source

Sources & research credit

1 source

Research by Gal Weizman · Forever Security

  1. EMBRACE: How We Hacked the Browsers Embedded in Top Five Desktop AI Assistants forever.security

Product behavior, dates, and patch status are from the Forever Security post. The post points to a technical companion for reproduction steps. Those steps are not repeated here. Link forever-agents-md-hijack-2026 and thisbash-claude-dispatch-rce-2026-11, which are different reports.

Record & downloads
Source record name
Forever Security EMBRACE, browsers embedded in five AI assistants
Entry ID
forever-embrace-embedded-browsers-2026-10
Catalog review
November 15, 2026
Category
attacks-on-ai
AI involvement confidence
confirmed