Threat reportVulnerabilityTL-2026-0873
AutoJack: AutoGen Studio MCP WebSocket Exploit Chain Turns an AI Browsing Agent into a Host RCE Vector
Microsoft Defender Security Research disclosed AutoJack, a three-bug exploit chain in AutoGen Studio's MCP WebSocket that lets a single malicious web page, when rendered by a local AI browsing agent, reach the agent's localhost control channel and spawn arbitrary commands on the host.
Key facts
| Identifiers | TL-2026-0873 |
|---|---|
| Also known as | AutoJack |
| Severity | HIGH |
| Affected products | Microsoft AutoGen Studio (autogenstudio); Microsoft AutoGen Studio (PyPI stable) |
| Weaknesses (CWE) | CWE-1385, CWE-306, CWE-78 |
| Tools | Web Content Summarizer (AutoGen Studio agent) |
| MITRE ATT&CK techniques | T1036 Masquerading, T1059 Command and Scripting Interpreter, T1071 Application Layer Protocol, T1106 Native API, T1132 Data Encoding, T1189 Drive-by Compromise, T1190 Exploit Public-Facing Application, T1199 Trusted Relationship, T1203 Exploitation for Client Execution, T1211 Exploitation for Stealth, T1559 Inter-Process Communication, T1571 Non-Standard Port and 4 more in the attack flow |
| Motivation | UNKNOWN |
| Target sectors | technology, software development, ai/ml research |
| Target regions | Global |
| Detection rules | 9 (1 shown free) |
| Sources | 12 cited sources |
Detection coverage
As of 2026-06-19, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-0873 across Splunk SPL, Microsoft KQL and Sigma, covering 21 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.
Free detection: AutoJack Exploitation — MCP WebSocket to localhost:8081 carrying server_params (SPL)
Detects the AutoJack entry point: a WebSocket/HTTP upgrade request to the AutoGen Studio MCP control channel /api/mcp/ws/ on localhost:8081 that carries a base64 server_params query parameter. This is the chained CWE-1385 (localhost origin allowlist defeated by the on-host browsing agent) + CWE-306 (AuthMiddleware skip-list) + CWE-78 (server_params decoded into StdioServerParams and spawned) entry vector. Post-patch flows use POST /api/mcp/ws/connect with a one-time session ID and never carry inline server_params, so an inline server_params on the legacy path is a strong indicator.
- Languages: Splunk SPL, Microsoft KQL, Sigma
- ATT&CK: T1190, T1189, T1659, T1132.001
- Data sources: proxy, stream:http, DeviceNetworkEvents, suricata
- Severity: high
- Confidence: high
Splunk SPL
index=proxy OR index=web sourcetype IN ("stream:http","pan:traffic","cisco:wsa:w3c","suricata")
| eval uri=coalesce(uri_path,url,cs_uri_stem,c_uri)
| eval q=coalesce(uri_query,url,cs_uri_query,c_uri_query)
| search (dest_port=8081 OR dest="127.0.0.1" OR dest="localhost" OR dest="::1")
| search uri="*/api/mcp/ws/*" AND q="*server_params=*"
| where NOT match(uri, "/api/mcp/ws/connect")
| rex field=q "server_params=(?<b64_params>[A-Za-z0-9+/=%]+)"
| stats count min(_time) as first_seen max(_time) as last_seen values(uri) as uri values(src) as src by dest dest_port
| eval reason="AutoJack: inline base64 server_params on legacy MCP WebSocket path (CWE-1385/306/78)"Microsoft KQL
// Sentinel: MCP WebSocket entry point carrying server_params to AutoGen Studio
let lookback = 24h;
DeviceNetworkEvents
| where Timestamp > ago(lookback)
| where RemotePort == 8081 or RemoteUrl has "localhost:8081" or RemoteUrl has "127.0.0.1:8081"
| where RemoteUrl has "/api/mcp/ws/" and RemoteUrl has "server_params="
| where RemoteUrl !has "/api/mcp/ws/connect"
| project Timestamp, DeviceName, InitiatingProcessFileName, InitiatingProcessCommandLine, RemoteUrl, RemoteIP, RemotePort
| extend Reason = "AutoJack inline server_params on legacy MCP WebSocket (CWE-1385/306/78)"Sigma
title: AutoJack MCP WebSocket Request Carrying server_params to localhost:8081
id: f1a0c873-1001-4a01-9b10-autojackspl001
status: experimental
description: Detects WebSocket/HTTP requests to the AutoGen Studio MCP control channel /api/mcp/ws/ on port 8081 carrying an inline base64 server_params parameter (AutoJack CWE-1385/CWE-306/CWE-78 entry point).
references:
- https://www.microsoft.com/en-us/security/blog/2026/06/18/autojack-single-page-rce-host-running-ai-agent/
- https://github.com/microsoft/autogen/pull/7362
author: Threadlinqs Team
date: 2026/06/19
tags:
- attack.initial_access
- attack.t1190
- attack.t1189
- attack.t1659
- cwe.2026.1385
logsource:
category: proxy
detection:
selection:
c-uri|contains:
- '/api/mcp/ws/'
c-uri|contains: 'server_params='
selection_port:
dest_port: 8081
filter_patched:
c-uri|contains: '/api/mcp/ws/connect'
condition: selection and selection_port and not filter_patched
falsepositives:
- Post-patch /api/mcp/ws/connect session flows (no inline server_params)
level: highKnown false positives:
- Post-patch POST /api/mcp/ws/connect session-ID flows that carry no inline server_params
- Developer/QA testing of the MCP WebSocket route in a controlled lab
Tuning notes: Scope to known AutoGen Studio developer workstations and the default port 8081 (adjust if bound elsewhere). Exclude the post-patch /api/mcp/ws/connect path. If proxy logs lack the query string, pivot to host network telemetry (Sysmon EID 3 / DeviceNetworkEvents) for connections to 8081 from python/headless-browser processes.
- AutoJack Exploitation — Unauthenticated MCP WebSocket Accept with server_params (KQL) · unlock with TL Creds
- AutoJack Post-Exploitation — AutoGen Runtime Spawns Interpreter with pwned Env Marker (KQL) · unlock with TL Creds
- AutoJack Delivery — MultimodalWebSurfer Headless Browse to Untrusted Domain (KQL) · unlock with TL Creds
- AutoJack Exploitation — base64 StdioServerParams on MCP WebSocket (Sigma) · unlock with TL Creds
- AutoJack Post-Exploitation — Interpreter Spawn from AutoGen Runtime on Linux (Sigma) · unlock with TL Creds
- AutoJack Auth/Origin Bypass — Unauthenticated /api/mcp WebSocket Accept (Sigma) · unlock with TL Creds
- AutoJack Post-Exploitation — Anomalous Child Process of AutoGen Studio Runtime (SPL) · unlock with TL Creds
- AutoJack C2/Delivery — Browsing Agent to Untrusted Domain then Localhost MCP Callback (SPL) · unlock with TL Creds
Indicators of compromise
As of 2026-06-19, Threadlinqs Intelligence tracks 21 indicators of compromise for TL-2026-0873.
Technical analysis
The chain combines a localhost-trusting origin allowlist (CWE-1385), an authentication middleware that opts MCP paths out of auth (CWE-306), and a command endpoint that runs an executable taken straight from a URL parameter (CWE-78). Demonstrated as a proof of concept (calc.exe pop); no exploitation in the wild and no CVE assigned.
AutoJack is an exploit chain disclosed by Microsoft on June 18, 2026 that demonstrates how an AI browsing agent can be coerced into becoming the attacker's last-mile remote-code-execution delivery vehicle. The target is AutoGen Studio, the low-code UI for Microsoft Research's AutoGen multi-agent framework. AutoGen Studio runs a local web service (default http://localhost:8081) and, in affected development builds on the main branch, exposed a Model Context Protocol (MCP) WebSocket control channel under /api/mcp/ws/.
The attack chains three independent weaknesses. First (CWE-1385, Missing Origin Validation), the MCP WebSocket enforced an origin allowlist that trusted http://127.0.0.1 and http://localhost. That check is meant to stop an ordinary remote browser pointed at a malicious site, but an AutoGen browsing agent (MultimodalWebSurfer) runs a headless browser on the same host, so any page it loads inherits the localhost identity and satisfies the origin check. Second (CWE-306, Missing Authentication), the AuthMiddleware contained an early-return skip-list for WebSocket-style paths including /api/ws and /api/mcp, on the assumption that the MCP handler would verify tokens itself; the handler never implemented that follow-up check, so the socket accepted unauthenticated connections regardless of the configured authentication mode. Third (CWE-78, OS Command Injection), the WebSocket endpoint accepted a base64-encoded server_params query parameter, decoded it into a StdioServerParams object, and passed the resulting command and args directly to stdio_client(...) with no allowlist on which executable could launch.
The end-to-end flow: a developer runs AutoGen Studio with a browsing agent on localhost:8081; an attacker plants a malicious page on the internet (or injects a URL into the agent prompt); the agent's MultimodalWebSurfer navigates a headless browser to the attacker page; the page's JavaScript opens ws://localhost:8081/api/mcp/ws/?server_params=<base64>; AutoGen Studio decodes the base64 JSON into StdioServerParams and spawns the requested command (e.g., calc.exe or powershell.exe) under the developer's account. Microsoft's proof of concept used a 'Web Content Summarizer' agent that, when fed an attacker URL, pops calc.exe on the developer's desktop.
Microsoft reported the behavior to the Microsoft Security Response Center (MSRC); the maintainers hardened the upstream main branch in commit b047730 (PR #7362). The fix stops reading the command from the URL: parameters are now stored server-side behind a one-time session ID via a new POST /api/mcp/ws/connect route, unknown IDs are refused, and MCP routes now flow through the normal authentication path (the middleware no longer exempts /api/mcp). Microsoft states the vulnerable MCP WebSocket surface was never included in a PyPI release, so the current PyPI build (0.4.2.2, which does not include autogenstudio/web/routes/mcp.py) is not exposed to this specific chain. The originating hunt and The Hacker News reporting note the dev pre-releases 0.4.3.dev1 and 0.4.3.dev2 on PyPI and that neither build was yanked; sources differ on whether the vulnerable route reached a packaged pre-release, so source installs and dev builds tracking main during the affected window should be treated as the at-risk population. Microsoft emphasizes the broader lesson: when an agent on your machine can browse the open web and talk to privileged local services, localhost stops being a trust boundary; the pattern is broader than this single bug.
MITRE ATT&CK techniques by tactic:
Defense Evasion
T1036 Masquerading; T1211 Exploitation for Stealth
Execution
T1059 Command and Scripting Interpreter; T1106 Native API; T1203 Exploitation for Client Execution; T1559 Inter-Process Communication
Command and Control
T1071 Application Layer Protocol; T1132 Data Encoding; T1571 Non-Standard Port
Initial Access
T1189 Drive-by Compromise; T1190 Exploit Public-Facing Application; T1199 Trusted Relationship; T1659 Content Injection
Resource Development
T1583 Acquire Infrastructure; T1587 Develop Capabilities; T1608 Stage Capabilities
Remediation
Patches
- Apply upstream AutoGen Studio fix commit b047730 (PR #7362)
Immediate actions
- Inventory hosts running AutoGen Studio from source or dev builds (0.4.3.dev1/0.4.3.dev2 or any main-branch checkout) that include autogenstudio/web/routes/mcp.py
- Update to a build at or after upstream commit b047730 (PR #7362) where MCP parameters are bound server-side and MCP routes flow through normal auth
- If immediate patching is not possible, do not run AutoGen Studio browsing agents (MultimodalWebSurfer) against untrusted URLs or untrusted prompt input
- Bind the AutoGen Studio service to a restricted interface and firewall localhost:8081 from agent-driven headless browsers where feasible
Workarounds
- Disable or remove the MCP WebSocket route (autogenstudio/web/routes/mcp.py) in affected source installs
- Restrict the AutoGen Studio browsing agent from navigating to untrusted/non-corporate domains
- Stay on the current stable PyPI release (0.4.2.2), which does not ship the vulnerable MCP route
Longer-term hardening
- Treat localhost as an untrusted boundary on any host where an AI agent can both browse the open web and reach privileged local services
- Require authentication and origin validation on all local agent control channels, including WebSocket/MCP endpoints
- Allowlist which executables an agent control channel may launch; never pass user/URL-derived parameters directly to a process spawner
- Run browsing agents in isolated sandboxes or containers with no access to host-local control services
- Deploy EDR with behavioral detection for unexpected child processes of agent runtimes
Affected and fixed versions
- Microsoft — AutoGen Studio (autogenstudio)
Vulnerable versions: main-branch development builds including autogenstudio/web/routes/mcp.py (between MCP plugin landing and hardening); 0.4.3.dev1; 0.4.3.dev2
Fixed in: builds at or after commit b047730 (PR #7362) - Microsoft — AutoGen Studio (PyPI stable)
Fixed in: 0.4.2.2 (does not include the MCP WebSocket route; not exposed to this chain)
Timeline
- Microsoft publishes advanced-hunting guidance to detect WebSocket connections to localhost:8081 carrying a server_params parameter and anomalous child processes of the AutoGen Studio runtime.
- Microsoft demonstrates a proof of concept: a 'Web Content Summarizer' browsing agent fed an attacker URL pops calc.exe on the developer's desktop; no in-the-wild exploitation reported.
- Upstream fix landed in commit b047730 (PR #7362): server-side parameter binding behind a one-time session ID, refusal of unknown IDs, and MCP routes flowing through normal authentication.
- Behavior reported to the Microsoft Security Response Center (MSRC); maintainers hardened the upstream AutoGen Studio main branch prior/concurrent to disclosure.
- Microsoft Defender Security Research (Shaked Ilan, with contributions from Microsoft Threat Intelligence) publishes the AutoJack disclosure detailing the AutoGen Studio MCP WebSocket exploit chain.
- Sources note PyPI stable 0.4.2.2 does not ship locked-ioc-3 and is unaffected, while source/main-branch checkouts and dev pre-releases 0.4.3.dev1/0.4.3.dev2 (not yanked) constitute the at-risk population.
- Defensive guidance issued: patch to commit b047730, isolate browsing agents from local control channels, and treat localhost as an untrusted boundary on agent-capable hosts.
- Assessed HIGH severity, POC_PUBLIC exploitability; current PyPI stable 0.4.2.2 not exposed (does not ship the MCP route), source/dev-build population treated as at-risk.
- Threadlinqs Intelligence opens tracking TL-2026-0873 on PoC-availability and developer-workstation / local-service-exposure risk; no CVE assigned and no CVSS published in sources.
- The Hacker News and other outlets report the AutoJack chain, noting dev pre-releases 0.4.3.dev1/0.4.3.dev2 and that neither build was yanked.
Sources cited
- AutoJack: How a single page can RCE the host running your AI agent Microsoft · Vendor report · 2026-06-18
- CWE-1385: Missing Origin Validation in WebSockets MITRE · Research · 2026-06-19
- CWE-306: Missing Authentication for Critical Function MITRE · Research · 2026-06-19
- CWE-78: Improper Neutralization of Special Elements used in an OS Command (OS Command Injection) MITRE · Research · 2026-06-19
- AutoGen Studio on PyPI (stable 0.4.2.2) pypi.org · Vendor report · 2026-06-19
- AutoGen Studio fix: bind MCP server_params server-side behind a one-time session ID (PR #7362) GitHub · Code repository · 2026-06-18
- AutoGen Studio hardening commit b047730 (MCP routes flow through normal auth) GitHub · Code repository · 2026-06-18
- Microsoft AutoGen — multi-agent framework repository GitHub · Code repository · 2026-06-18
- AutoJack Attack Lets One Web Page Turn an AI Browsing Agent into an RCE Vector The Hacker News · News · 2026-06-19
- Microsoft says web-enabled AI agents can trigger host-level RCE csoonline.com · News · 2026-06-18
- Microsoft warns AI agents are being 'AutoJack'-ed to deliver RCE payloads by browsing untrusted websites techradar.com · News · 2026-06-18
- Microsoft Discloses AutoJack: A Malicious Webpage Can Hijack AutoGen Studio via Localhost RCE windowsnews.ai · News · 2026-06-18
Previous and next in vulnerability
- CVE-2026-54420 — LiteSpeed cPanel Plugin Symlink-Following (CWE-61) Privilege Escalation to Root on CloudLinux/CageFS Shared Hosting; Added to CISA KEV After In-the-Wild Exploitation Previous report
- CVE-2026-8713: Avada (Fusion) Builder WordPress Plugin Unauthenticated Path Traversal Arbitrary File Deletion Next report
Drafted by the Threadlinqs research pipeline from 12 sources; published ; last updated —.