Backed by $12 million in seed funding, Metano helps enterprises secure AI agents wherever they operate, including endpoints, cloud and SaaS.
Metano Product Team
A compromised AI tool could slip an ordinary-looking link into Jan’s chat. A single click replaced the app’s own window with an attacker page, handed it the app’s internal powers, and let it run any program on the user’s machine. No password, no second prompt.
| Product | Jan (janhq/jan) |
|---|---|
| Class | One-click RCE via untrusted content |
| Affected | <= 0.8.7 |
| Fixed in | 0.8.8 |
| Severity | Critical |
| CVSS 3.1 (assessed) | 9.6 |
| Advisory | GHSA-j6hv-vhmg-vpj6 |
| CVE | pending |
| CWE | CWE-94 / CWE-346 / CWE-1188 |
Jan is a widely used open-source desktop app for running AI models offline. It can connect to MCP servers, external tools the model calls to do real work. A malicious or compromised MCP server (or prompt-injected tool output the model relays) could put a normal markdown link into the chat. Jan rendered that link as a live, same-frame link, had no guard stopping the main window from navigating away, and shipped without a permissions manifest for its internal commands.
Chained together, one click navigated Jan’s main window to an attacker-controlled page, which then inherited the app’s internal bridge and called the command that spawns processes, achieving remote code execution as the user. Fixed in v0.8.8; all versions <= 0.8.7 are affected.
AI desktop apps have quietly become some of the most powerful programs on a user’s machine. They read files, run tools, and increasingly take actions through MCP, the standard that lets a model call external tools. That power comes with a new trust boundary that classic app security never had to think about: content produced by a model or a third-party tool is now rendered directly into a trusted, privileged window.
The industry has already seen where this leads. Prompt-injection chains like EchoLeak (CVE-2025-32711) turned a crafted email into silent data theft from an AI assistant. This finding is the next rung on the same ladder: not just data exfiltration, but full code execution, reached through the ordinary MCP tool feature working exactly as designed. The lesson generalizes to nearly every AI app built on a web stack (Electron or Tauri): if an untrusted model or tool output can steer a privileged window, a chat message becomes an exploit primitive.
Jan is an open-source, offline alternative to ChatGPT that runs models locally on your own machine. It is built by Menlo Research on the Tauri framework (a Rust backend with a web-based UI), and it is widely adopted.
Two design facts set the stage. First, MCP servers are semi-trusted at best. Users add them to extend the model with tools, and those tools (or the model relaying their output) can return arbitrary text into the chat. Second, Tauri apps talk between the web UI and the Rust backend over an IPC bridge: the frontend calls backend commands over it. Which origins may call which commands are meant to be governed by a permissions manifest (an ACL). Jan shipped without one, which turns out to matter a great deal.
IPC (inter-process communication): The private channel the app’s visible UI (a web page) uses to ask its powerful backend to do real work, like reading a file or starting a program. Think of it as an intercom between the front desk and the back office. It is meant to be used only by the app’s own pages.
ACL (access control list) and the ACL bypass: An ACL is the guest list for that intercom: the rules saying which pages are allowed to request which backend actions. An ACL bypass is when the guest list fails to stop someone it should have. Here the app shipped with no guest list at all for its own commands, so in the pinned framework version any page, including the attacker’s, could pick up the intercom and order the back office to launch a program.
No single bug here is fatal on its own. The severity comes from three independent weaknesses lining up so that untrusted chat content reaches the one backend command that can start a process.
[open report](http://attacker.tld) gets a link that navigates Jan’s own window.if command belongs to a plugin OR app has a manifest:
require a matching ACL entry # remote page rejected
else:
allow # Jan's case: any originTauri closed this in 2.11.1 (PR #15266); the changelog says custom commands “bypassed ACL entirely without an AppManifest, allowing any origin to invoke them.”
5. A backend command spawns an arbitrary process. The backend command that starts an MCP server takes the program, arguments, and environment straight from its caller and launches them, with no allowlist. Called from the attacker’s page, that is code execution as the user.

Figure 1: The attack chain. Break any one link in this chain and the attack fails.
The entire exploit is a link plus a tiny page. When the user clicks, the attacker’s page runs one line against the app’s internal bridge. See the full reproduction in the published advisory GHSA-j6hv-vhmg-vpj6. The payload used for the demo is a harmless touch /tmp/jan_pwned to prove execution; substitute any command and it runs with the user’s privileges. After the single click there is no prompt and no confirmation. The attack works from any origin because app commands skip the ACL entirely; it does not depend on the app’s capability allowlist. All testing was local and non-destructive against the open-source build.
Remote code execution as the user is the ceiling, the most severe outcome a vulnerability can have. Once a process spawns, the attacker can read and exfiltrate local files (including the user’s model data, API keys, and credentials), install persistence, pivot to other systems, or deploy ransomware, all with the victim’s privileges.
Any Jan user who connects the app to a third-party MCP server, which is the intended way to use the feature. The malicious content does not require the user to visit a shady website: the model relaying a poisoned tool result can deliver the link inside a normal-looking answer, so even a careful user who only ever talks to their own assistant can be hit if one connected tool is compromised or prompt-injected.
The one precondition is a single user click on a link in the chat, a routine action. Everything else is the default build: no special configuration, no non-default toggle. Affected versions are <= 0.8.7 across all platforms Jan ships.
This is a defense-in-depth collapse, not one careless line. A privileged desktop app should assume that model and tool output is hostile and keep several independent barriers between that content and any dangerous capability. Jan had three natural barriers, and all three were open at once: link rendering did not neutralize untrusted links, navigation was not constrained to the app’s own origin, and the IPC layer had no permissions manifest to deny remote callers. The upstream Tauri framework contributed the sharpest edge: in the pinned version, custom app commands with no manifest were callable by any origin, a primitive the Tauri team themselves later closed. Each barrier alone would have broken the chain; the finding exists because none were present on the path the content actually took.
Jan fixed this in v0.8.8. The chain breaks at any link, so a robust fix closes several:
On an unpatched build, signs of exploitation include the Jan main window navigating to an unexpected external origin, and child processes spawned by Jan that do not correspond to a configured MCP server binary. Monitoring process creation with Jan as the parent is the clearest signal.

Parth Shukla
AI Researcher
Parth is an Offensive Security Engineer at Metano. He breaks the AI ecosystem, then builds what keeps it from breaking.
The latest news, technologies, and resources from our team.
Backed by $12 million in seed funding, Metano helps enterprises secure AI agents wherever they operate, including endpoints, cloud and SaaS.
Metano Product Team
A prompt-injected agent in Amazon’s Kiro Crew could disable its own sandbox by editing a single config file. The file was writable from inside the sandbox, and a feature that skipped the security gate removed the last thing that would have questioned it.
Harsh Akshit
Since 21 October, computers infected with Mini Shai-Hulud have uploaded stolen secrets to 990 public GitHub repositories; the newest was created just 12 minutes before we counted. They upload using a stolen GitHub token that infected computers pass among themselves.
Japneet Singh
Anthropic’s inference hooks let a security server allow or deny every Claude Enterprise prompt and tool call before the model runs. Metano is now that server. It works on any device, with nothing installed.
Metano Product Team
Exposed LiteLLM gateways that still use the sample admin key will hand out API keys to anyone who asks. Each key is unique, so we could follow it to see who used it next.
Neelima Rustagi and Sonali Tyagi
An agent skill runs with everything the agent that loads it can do. New research from Metano MARS Labs, Purdue University and VulNow profiles more than 6,000 skills sourced from a couple of different marketplaces, to measure how much of that authority they put into play.
Japneet Singh
Occasional updates, new content, and insights. No spam; unsubscribe anytime.