BlogVendor review

Manus security review — prompt injection meets your logins

Manus reads the open web while signed into your accounts. Four published prompt-injection chains show why the approval depends on what it holds.

Drel Research11 min read

Manus is a general-purpose autonomous agent. A user gives it a goal and it plans the work, browses the web, writes and runs code in a cloud machine, reads documents and email through connectors, and hands back a result. The question a security team gets is short: can we approve it? The useful answer depends less on Manus's certificates than on which accounts the agent will be signed into while it reads content nobody in the organization wrote.

This review is built only on public information, as of November 2026: Manus's security page, documentation, blog and help center, the researchers' own write-ups, and press reporting. Drel has not tested or audited Manus. Where something is not publicly documented, it is listed as an evidence gap and a question for the vendor.

The short answers

Does Manus have SOC 2? Manus's security page lists SOC 2 Type 1 and SOC 2 Type 2 with letters of attestation, ISO 27001:2022 with a certificate, and ISO 27701:2019, and links to a trust center. It also lists control statements such as penetration testing performed, data encryption used, and customer data deleted on leaving. The page does not state the scope or period of the reports.

Does Manus train on customer data? Not confirmed either way from the sources we could read. The security page does not address training. Search engines still index a Manus help-center article on the question, summarized as saying Manus uses only aggregated or de-identified information to improve its services, but the article returned “not found” when we opened it, it is not in the current privacy and policies collection, and we could not retrieve the privacy policy text. Ask for the answer in writing, including whether an opt-out exists.

Who owns Manus, and where is the data? This changed twice in a year. Manus's own note to users (11 August 2026) says Meta acquired Manus on 29 December 2025 and that Manus will soon return to operating as an independent company. The Next Web reported in April 2026 that China's National Development and Reform Commission had ordered the deal canceled, and described Manus as incorporated in Singapore. The note says data is stored in the US and Singapore, and that data generated by certain users on or after 29 December 2025 would be deleted on 23 and 24 August 2026 to comply with regulatory requirements, not because of a security incident. We found no public statement that the separation has completed.

Retention and admin controls. The sandbox post says an inactive sandbox may be recycled after 7 days on the Free plan and 21 days on Manus Pro. A help article (1 October 2026) says content is deleted when an account is deleted, with some personal information kept for legal reasons, and that on Team plans retention may be set by the owner or administrator. The security page links to single sign-on for Team subscriptions. Other admin controls were not publicly documented in what we read.

What Manus can access and do

Four surfaces, each with a different blast radius:

  • A cloud sandbox per task. Manus describes a fully isolated cloud virtual machine with networking, a file system, a browser and software tools, in which the agent has full control including root access.
  • A cloud browser. The cloud browser docs say a user can log into their own accounts in it so Manus can act on them, and that those authenticated sessions are reused for later tasks.
  • Connectors. Linked services such as email. Manus says connectors are automatically disabled when collaboration is enabled on a task.
  • The Browser Operator extension. The Browser Operator docs describe an extension for Chrome and Edge that lets Manus work in the user's local browser, with its existing logins and active sessions.

This is a wider mandate than a coding agent has. The Devin review deals with an agent whose world is mostly a repository and a pipeline. Manus is pointed at the open web, an inbox and whatever the browser is signed into, which is why the agentic AI security review questions about tool scope and untrusted input apply in full.

Sandbox, browser and credentials

The sandbox

Manus's sandbox post (January 2026) says each sandbox runs in its own environment, that actions inside it affect only that sandbox, that session and account data cannot be accessed from it, and that the model follows a Zero Trust principle. It also says what a sandbox holds: uploaded attachments, files created during the task, and task configuration including tokens the user uploads and tokens Manus assigns for calling APIs. Sharing a task exposes messages and output artifacts only; collaboration lets invited users instruct the agent and, through it, read or change sandbox files. Manus advises checking a sandbox for sensitive content before adding collaborators.

Isolation between sandboxes limits what one task can do to another. It does not limit what an agent that has been steered by injected text can do with the files, tokens and network access inside its own sandbox. That distinction is the subject of agent sandboxing and isolation.

Logged-in sessions

For both browsers Manus states that it does not store passwords. In the cloud browser, the user completes multi-factor prompts through a “Take Over” step and can sign out of accounts or clear all sessions in settings; the docs say sessions are encrypted and each user gets a separate browser instance. With the Browser Operator, the user authorizes each session, the agent works in a dedicated tab, and actions are logged. A password is not the asset here. A signed-in session is, and the agent holds it for the length of the task.

Prompt injection: the published research

Four independent groups have published work showing Manus following instructions hidden in content it was asked to read. No CVE IDs were assigned in the write-ups we read. The table gives each source's own account; we have not reproduced any of it.

Published research, as described by each source

SourceDatesWhat was shownFix status
Embrace The RedReported 1 Jun 2025, published 25 Aug 2025Hidden instructions in a PDF made the agent expose its internal VS Code Server to the internet and send the URL and password to an attacker's site.The post says the status of specific mitigations remained unclear at publication.
Aurascape AuraLabs (SilentBridge)Reported 18 Sep 2025, published 24 Feb 2026Three zero-click variants (web page, search result, document), leading to Gmail data theft through the connector, secrets in the sandbox, and a root shell in the container.The post says Manus deployed mitigations for the critical issues in November 2025.
CodeAnt AIReported 10 May 2026A shared project's owner-set instruction ran a remote script in an invited user's sandbox, leaking a static desktop password and giving control of that user's remote desktop.The post says it was triaged and fixed on 16 Sep 2026.
Salt LabsReported in the press 24 Sep 2026An email the agent read carried obfuscated JavaScript that ran before the safety warning appeared; a reverse shell then retrieved tokens for linked accounts.Reported as patched. We did not read Salt's own write-up.

Embrace The Red showed the first chain: while summarizing a PDF, the agent called its port-exposure tool without asking the user, read the VS Code Server password from a local configuration file, and passed both to an attacker's site with its browsing tool. The same post notes that Manus rendered images from untrusted domains, a second route for sending data out.

Aurascape's SilentBridge research rated three variants CVSS 9.8 by its own scoring. A request as ordinary as “summarize this page” led the agent to query a connected Gmail account and send the results to an attacker. A document variant opened a reverse shell in the sandbox, where passwordless sudo gave root in the container. The post also reports API keys in container variables and access to the instance metadata address and internal network hosts. It says Manus acknowledged the report on 5 October 2025 and deployed mitigations in November 2025.

CodeAnt AI described a stored variant: the owner of a shared project sets an instruction, an invited user runs any task, and the agent fetches and runs a script that leaks sandbox variables, including a static password for the sandbox's remote desktop. With that, the attacker could reach the browser sessions the invited user had open. The post says it was reported to Meta's bug bounty program on 10 May 2026, closed on 27 August, resubmitted on 4 September, and fixed on 16 September 2026. The Salt Labs finding is known to us only through press coverage of a Dark Reading report.

Read together, the reports do not describe one bug that was fixed. They describe a property of the design: the agent reads untrusted content, holds signed-in sessions and tokens, and can send data out. Each fix closes a route. The review has to decide what the agent may hold while that property remains.

Manus's help center carries an article titled “Verify before you act”, and its docs recommend monitoring multi-step workflows. That places part of the control on the user. For the mechanics of the attack class, see prompt injection explained.

Model governance and sub-processors

The security page names no model providers and no sub-processors; it links to a models page and to the trust center, which we could not read without a browser session. For a product that forwards the content of mailboxes and logged-in pages to language models, this is the largest documentation gap. A review needs, in writing: which providers process task content, in which regions, whether they retain it, whether any of them or Manus may train on it, and how customers are told when a provider or model changes.

The CodeAnt write-up also reports a live LLM proxy key among the variables readable inside a sandbox. Whether that key was scoped to one user or one task is not stated, and is a fair question for the vendor.

Compliance and ownership

On paper the certification set is broad for a company of this age: SOC 2 Type 1 and Type 2, ISO 27001:2022 and ISO 27701:2019. Two things limit how much weight it carries in a review. The reports describe organizational controls and do not test whether an agent can be steered by a web page, the gap covered in SOC 2 is not AI assurance. And the audit periods ran across a change of ownership that is, by Manus's own account, being reversed.

Ownership matters here for practical reasons, not political ones. It decides who the contracting entity is, which law governs the data, where the infrastructure sits, and who runs the vulnerability intake; the 2026 research above went through Meta's bug bounty program. The August 2026 deletion shows that a regulatory event can remove customer data on about two weeks' notice. Any approval should name ownership change as a trigger for re-review.

Evidence gaps

These were not publicly documented in the sources we read, or could not be confirmed from them. Each is a question for the vendor:

  • Is customer content (prompts, uploads, task outputs, browser sessions) used to train any model, and is there an opt-out per user or per team? We could not retrieve the privacy policy text, and the security page does not address training.
  • Which model providers and sub-processors receive task content, in which regions, and under what retention terms? The security page names none.
  • Which legal entity is the contracting party and data controller today, who owns it, and has the separation from Meta completed?
  • What is the scope and audit period of the SOC 2 Type 2 report, and does it cover the sandbox, the cloud browser and the Browser Operator extension?
  • How are logged-in cloud browser sessions stored, for how long, and who at the vendor can access them? The docs say sessions are encrypted and passwords are not stored; they do not describe session storage.
  • Can an admin restrict sandbox network egress, disable connectors or the Browser Operator for a team, or require approval before the agent exposes a port or sends data to a new domain?
  • What audit log does a team owner get: every tool call, every site visited, every connector read?
  • For each published finding, what was changed, and is there a public security advisory or disclosure channel? No CVE IDs were assigned in the write-ups we read.
  • What data did the August 2026 deletion cover, which users were affected, and what replaced the previous hosting arrangement?

Review checklist

An AI security review of a Manus rollout should be able to answer:

  • Which Manus surfaces are in scope: the cloud sandbox only, the cloud browser with logged-in accounts, connectors such as email and drive, the Browser Operator extension on employee laptops?
  • For each connector and logged-in account: what can that identity read and change, and is it a dedicated low-privilege account or a person's own?
  • What untrusted content will the agent read (web pages, search results, inbound email, uploaded documents, shared projects), and what can it reach in the same task?
  • Is anything irreversible (sending mail, payments, deleting records, publishing) possible without a person confirming it?
  • Are secrets, API tokens or customer files ever placed in a sandbox, and how long does that sandbox live before it is recycled?
  • Is the Browser Operator extension blocked or allowed by browser policy, and for which profiles?
  • Is single sign-on enforced on a Team plan, and who administers retention for the team?
  • What data classification is permitted in tasks, and is that written where users will see it?
  • Who is told, and what is revoked, if an agent task is found to have followed injected instructions?
  • What triggers a re-review: a change of ownership, a new connector, a new published finding?

Framing the decision

The decision follows from what the agent will hold, not from the product name. Written with its conditions, each outcome looks like this:

  • Approve for research and drafting on public information, in the cloud sandbox, with no connectors, no logged-in sessions and no confidential uploads.
  • Conditional approval where a connector or a logged-in account is needed: a dedicated low-privilege account, a named list of permitted services, confirmation by a person before anything irreversible, single sign-on on a Team plan, and the vendor's written answers on training and sub-processors. Each condition has an owner and a date.
  • Restricted pilot to learn how people use it: a small named group, synthetic or public data, the Browser Operator extension blocked, a fixed end date, and a list of what the pilot must show. See the restricted pilot pattern.
  • Reject, for now, for use that puts the agent inside a person's primary mailbox, an administrator session or systems holding regulated data, while the gaps on training, sub-processors and ownership stay open.

Whichever outcome is chosen, record what it rested on: the sources read, their dates, the open questions, and who signed. The facts in this review moved several times during 2026 and will move again.

Common questions

Is Manus AI safe to use?

Manus lists SOC 2 Type 2, ISO 27001:2022 and ISO 27701:2019 and runs each task in a separate cloud machine. Independent researchers have also shown, four times since 2025, that hidden instructions in content could make the agent leak data or run code; the sources report those specific issues as mitigated or fixed. Safety depends mostly on what accounts and data the agent is given.

Is Manus SOC 2 compliant?

Its security page lists SOC 2 Type 1 and Type 2 letters of attestation, available through its trust center. The scope and audit period are not stated on the public page, so request the report.

Does Manus use my data to train models?

We could not confirm this from a primary source. A help-center article that search engines summarize as “only aggregated or de-identified information” was no longer available when we looked, and the security page is silent on it. Ask the vendor for a written statement that covers prompts, uploads and browser content.

Is Manus owned by Meta?

Meta acquired Manus on 29 December 2025. After a Chinese regulatory order reported in April 2026, Manus told users on 11 August 2026 that it would soon return to operating as an independent company. Confirm the current contracting entity before signing.

Sources

Blog

Get new posts in your inbox

AI security review, OWASP Agentic Top 10, ISO 42001 evidence, and what AI Committees actually need. No cadence promises — we publish when there's something worth reading.

Review a Manus rollout before it holds a login

Connectors, signed-in sessions, and what the agent reads from outside, mapped to an approval decision your AI Committee can sign.

A note on scope: Drel reviews assessed systems against documented architecture, configuration and intent. It does not ingest live telemetry from production environments. Dispositions reflect the assessed system at the time of review and the re-assessment triggers that govern when the disposition must be revisited.