When attackers bring AI agents: protecting the systems that matter

Updated

Reuters recently reported on research from the security firm NorForge: attackers used an open-source tool called Cairn to send swarms of autonomous AI agents against more than 1,200 online stores in the US, Canada, Australia and Japan. Some got in after about a day of trying. Each successful break-in cost the attacker around $50 in AI usage. Well-known companies were on the list, and in at least one case the way in was a website run by a supplier (Futurism’s summary).

At the same time, vendors are shipping urgent patches for the software that runs critical infrastructure. If you look after systems that matter, the advice starts in the obvious place: apply those patches, quickly.

But it is worth being honest about what changed. One person can now probe a thousand targets at once, around the clock, for the price of lunch. Attackers no longer need to pick their targets carefully. They can try everything and keep whatever works. Some attempts will land, however well you patch. What decides whether that becomes a bad week or a disaster is what the attacker can reach once they are in.

That is mostly a question of identity and access: whose credentials they can use, what those credentials can do, and how long it takes you to notice and shut the door.

A practical checklist

1. Patch on a clock, and know what you run

You cannot patch what you do not know you have. Keep a current list of your internet-facing systems, including the ones your suppliers run for you. Set a deadline for critical patches (days, not weeks), and track the exceptions in writing, with a name and a date next to each one.

2. Assume passwords will be tried, and some will be stolen

Automated attacks are very good at trying leaked passwords everywhere. Require multi-factor authentication for every account that can change something, and phishing-resistant passkeys for administrators. Retire shared admin passwords; if three people know a password, you cannot tell which of them used it.

3. Remove standing administrator access

An admin account that is always on is an admin account an attacker can always use. Keep privileged credentials in a vault. Have people check them out for a stated reason and a set time, and change the password once they hand it back. For everyday access, grant it for hours, not forever.

4. Treat machine accounts and AI agents as identities too

Service accounts, API keys and AI agents often have more access than any person, and nobody reviews it. Give each one an owner. Give it the narrowest access its job needs, with credentials that expire and rotate. When it stops being used, switch it off. Attackers love a forgotten service account, because nobody is watching it.

5. Review access with the facts in front of you

Access builds up quietly. Review it on a schedule, and give reviewers what they need to decide: when the access was last used, and how it compares with people in similar roles. In the cloud, look in particular for broad permissions that nobody uses. They do nothing for you and a great deal for an intruder.

6. Split the steps that move money or data

The person (or agent) who creates a payment should not be the one who approves it. The same goes for exporting customer data or changing who can sign in. Separation of duties is what turns one stolen account into an inconvenience instead of a loss.

7. Plan for the bad day before it comes

Decide now how you will get in if your single sign-on goes down, and require two people to approve that emergency access. Practise removing someone’s access, or an agent’s, in minutes rather than days. When someone leaves, their access should go the same day, including the tokens and agents they set up.

8. Keep a record you can trust

When something happens, the first questions are who did what, and when. Keep one audit trail for people and machines alike, and make it tamper-evident. If an intruder can quietly edit the log, the log cannot tell you what they did.

9. Hold your suppliers to the same standard

In at least one case in the NorForge report, the way in was a website a supplier ran for the company. Give suppliers only the access they need, time-limit it, and review it like your own.

How AtmanID helps

AtmanID does not replace patching, endpoint protection or a web application firewall. It covers the part of the problem that decides how far an attacker gets: identity and access, for people and for software, in one place.

  • Privileged access without standing admin rights. Vaulted credentials are checked out with a justification and an automatic time limit, revealed for a short window, flagged for rotation when they come back, and changed under dual control so one administrator cannot do it alone. Emergency access during an SSO outage needs a second administrator to approve it (privileged access).
  • Access for hours, not forever. People can request access for a few hours through the catalog, and it is removed when the time runs out. For AI agents, a zero-standing-privilege view lists every standing credential and which ones have no documented, time-bound exception (same guide).
  • Strong sign-in. Multi-factor authentication per role, with passkeys required for sensitive roles (MFA enrollment).
  • Agents with an owner and a short leash. Agents and other non-human identities are registered with an owner and rotation policies, receive scoped tokens for specific tools, and every tool call passes through the AtmanID gateway for scope, rate and data checks (MCP gateway, rotation policies).
  • Reviews and separation of duties. Certification campaigns give reviewers usage and peer context (certification campaigns), and separation-of-duties rules block or warn on conflicting access (SoD rules).
  • Leavers handled the same day. The leaver workflow removes a person’s access, deactivates the AI agents they own and flags their other machine identities as orphaned (joiners, movers, leavers).
  • One record of what happened. People and agents share one tamper-evident, hash-chained audit trail (audit chain).

If you do three things this week

  1. Apply the critical patches your vendors have issued, and write down anything you cannot patch yet, with an owner.
  2. Find every account with standing administrator access, human or machine, and cut that list down.
  3. Turn on multi-factor authentication for every administrator, and check that your emergency access needs two people.

Further reading: AI agent security, how to govern AI agents, what is a non-human identity?