Ravindra BagaleCourses & study guides Track your progress

Guides · Generative AI

Manus and Autopilot AI Agents: CERT-In Risks and Safe Practice (2026)

By , trainer LinkedIn Published: Last updated: 5 min read

Autopilot-style AI agents (including tools discussed under names like Manus and similar browser/computer agents) can click, type and open sites for you. In 2026, Indian reporting around CERT-In advisories and press coverage has warned that agentic tools can exceed intended scope on live systems. This guide teaches safe practice only: throwaway accounts, no production passwords, human approval before send/pay/delete. It does not teach attacks, bypasses, or how to evade security controls.

Suppose a demo video shows an agent booking and paying alone. Impressive — and dangerous on your real bank login. Use a practice account, approve every money step yourself, and stop if the agent opens a strange site. Safety is the skill employers actually need.

Related

See also AI browser agents and deepfake scam guides on this site. This page is the agent autopilot risk landing for 2026 India awareness.

What you will learn in this guide

  • What autopilot agents are (plain words)
  • Why CERT-In style warnings matter
  • Never give production passwords
  • Throwaway test account demos
  • Human approve before send/pay/delete
  • Incident basics if something goes wrong

1. Suppose an agent "helps" by opening your net banking

Suppose a learner pastes a bank URL and saved password into an agent "so it can pay electricity". That is how accounts get drained or locked. The safe demo uses a fake shop admin on a test site, not real money rails.

2. Words

Word Simple meaning
Autopilot agent AI that takes actions (click, type, browse), not only chat
Scope What you allowed the agent to do
Exceeded scope Agent did more than you intended
Human approval You confirm before send, pay, or delete
Throwaway account Practice login with no real money or customer data
Safe autopilot practice in five steps Steps from top to bottom: 1 Use a throwaway account, No real money rails; 2 Block money permissions, No bank passwords; 3 Watch every action, Stop on strange pages; 4 Approve sends yourself, Pay and delete too; 5 Stop and report if wrong, Change passwords cleanly. 1Use a throwaway accountNo real money rails 2Block money permissionsNo bank passwords 3Watch every actionStop on strange pages 4Approve sends yourselfPay and delete too 5Stop and report if wrongChange passwords cleanly

Figure 1. Safe autopilot practice in five steps: use a throwaway account, block money permissions, watch every action, approve sends yourself, then stop and report if something looks wrong.

3. Why India warnings matter (2026)

  1. Read reputable coverage of CERT-In / agentic-tool risks (Moneycontrol, BusinessLine and CERT-In channels as reported in Oct 2026 research notes). Details evolve — follow official CERT-In publications for current advisories.
  2. Risk themes to remember: unexpected actions on live systems, over-broad permissions, and phishing-like trust in "the agent said it is fine".
  3. Your takeaway is defensive: least privilege, monitoring, human gates.
  4. This page will not reproduce exploit steps.

4. Worked example 1: never give production passwords

  1. Write a personal rule: agents never receive bank, email root, AWS root, or GST portal passwords.
  2. If a tool asks for a password manager export, refuse.
  3. Use API keys with spend caps for coding agents — not owner keys.
  4. Prefer read-only demo data.
  5. Log out shared machines after demos.

Expected output: a written rule on your desk sticky note.

5. Worked example 2: throwaway test account demo

  1. Create a practice Gmail/shop admin that holds no customer PII and no saved UPI.
  2. Give the agent only that login inside a disposable browser profile.
  3. Task example: "Collect three public product titles from my test catalogue page into a Sheet."
  4. Watch the screen. If it navigates to a payment or password-change page, stop.
  5. Delete the throwaway session cookies when finished.

Expected output: a boring successful demo with zero money movement.

Ravindra Bagale's Tip

Students record demos on their real Amazon seller account. One wrong bulk edit hurts live listings. Always clone a test store.

6. Worked example 3: human approve before send/pay/delete

  1. Configure any agent setting that requires confirmation before send mail, pay, or delete files — turn those confirmations ON.
  2. Checklist before you click Approve:
- Is the recipient correct?
- Is the amount correct?
- Is this the right environment (test vs live)?
- Did I type the OTP myself on an official app — not into the agent chat?
  1. If unsure, Deny and do the step manually.
  2. Keep OTP and UPI PIN out of agent prompts forever.
  3. For office use, follow written company policy; if none exists, escalate to IT before enabling agents on real systems.

7. If something looks wrong

  1. Stop the agent; disconnect network if needed for that machine.
  2. Change passwords for any account the agent touched (from a clean device).
  3. Check bank/UPI SMS for unknown payments; contact bank channels you already trust.
  4. Report cyber fraud patterns via official paths such as 1930 / cybercrime.gov.in when appropriate.
  5. Tell your trainer or IT what permissions were granted — for learning, not for blame theatre.

8. Beginner mistakes and quick fixes

Mistake What happens Quick fix
Real bank login in agent Account risk Throwaway only
Auto-approve all actions Scope creep Require prompts
Ignoring CERT-In news Repeat incidents Read advisories
OTP typed into agent chat Takeover risk Official app only
Demo on prod seller account Live damage Test catalogue

9. Try at home

Try at home

  1. Write your three hard rules: no prod passwords, throwaway demos, human approve pay/send/delete.
  2. Run one agent task on a test account only.
  3. Practice hitting Stop when the agent opens an unexpected URL.

Recap

In short

  1. Autopilot agents act — they are not only chatbots.
  2. India 2026 warnings stress scope and live-system risk.
  3. Never hand production passwords or OTP to agents.
  4. Demo on throwaway accounts.
  5. Human approval before send, pay, or delete.

Impressive demos are cheap; recovering a drained account is not. Practice safe.

Frequently asked questions

Is Manus banned in India?

Follow current CERT-In and official advisories for any named product. This page teaches safe practice for the whole agent class, not a legal ban list.

Can I use agents at work?

Only with written permission, least privilege, and IT-approved tools.

Does this teach hacking?

No. Safe practice and awareness only.

What is the one rule to remember?

If money, mail send, or delete is involved, you approve it — not the agent alone.