Breaking

Salt Labs Says a Single Email Could Hijack Manus AI Agent Before a Guardrail Intervened

October 4, 2026 at 05:06 EDT

Security
AI Agents

Salt Labs disclosed research showing that Manus, an agentic AI platform, could be hijacked through one malicious email and made to run code inside a user's cloud sandbox before its warning system intervened. The issue has been responsibly disclosed, remediated, and is no longer exploitable, according to Salt Security.

Oct. 1, 2026 · Salt Labs

One Email, Agent-Wide Risk

A Manus test showed malicious email content could become executable action before a guardrail warning appeared. The flaw has been fixed, but the pattern matters for any agent connected to real accounts.

Email delivered payload

Events needed

36%

Users granting email access

Fixed

No longer exploitable

Where users are connecting agents

36%

33%

Browser

31%

Messaging

29%

Storage

27%

Calendar

Attack path

Email arrives

→

Agent reads inbox

→

Node.js runs payload

→

Warning appears late

Why agents are useful

They can read mail, summarize content and act across connected tools from natural-language instructions.

Why this is risky

If untrusted content reaches tools before policy checks, a warning can arrive after the agent has already acted.

The security lesson

Prompt inspection is necessary, but connected agents also need pre-execution controls over tools, APIs, tokens and outbound behavior.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.