Attacker agents and your own agents
Protecting a company from AI agents has two sides: the agents attackers point at you, and the agents you run yourself. The first group uses AI to find and exploit weaknesses faster and at greater scale. The second group is software you have given access to your data and tools, which anyone who can put instructions in front of it may be able to steer. Both do damage through the same thing: whatever they can reach.
Defence against attacker-run AI agents
Defending against attacker-run agents means shrinking what they can find and fixing what they find before they do.
- Inventory and cut internet exposure. List every public IP address, domain, login page and API. Take off anything that does not need to be public, and put admin tools behind your identity provider.
- Patch exposed systems within days. Google’s analysis put the average time from disclosure to exploitation at five days in 2023, before AI tools made scanning cheaper still.
- Use phishing-resistant MFA for every admin and every email account. CISA calls FIDO and WebAuthn authenticators the gold standard.
- Retire long-lived shared secrets. Every API key in a repository, a CI log or an app binary is something an automated scanner can find and use. What happens when an API key leaks? covers the clean-up.
- Watch for automation in your logs. Bursts of requests, odd paths and sign-in attempts at machine speed are the signature of automated attackers.
Can AI hack my website? goes deeper on the public web side.
AI agent governance
Govern each AI agent your company runs like a contractor with a badge, not like a feature.
- Keep an inventory. Every agent, what it is for, who owns it, which tool servers it uses and what data it touches. Shadow agents connected by individual employees are the ones nobody revokes.
- Give every agent its own identity. A named service identity per agent, so its actions are attributable and it can be cut off on its own.
- Grant the narrowest tools. OWASP’s guidance on excessive agency is to limit the extensions an agent can call to the minimum necessary and avoid open-ended ones such as a general command line.
- Use credentials that expire. AWS recommends temporary credentials for workloads instead of long-lived keys; the same applies to agents. Scope them to one task.
- Gate irreversible actions. OWASP recommends requiring a human to approve high-impact actions. Payments, deletions, permission changes and messages leaving the company qualify.
- Vet tool servers before installing them. The Model Context Protocol security guidance says clients must show the exact command and get explicit approval before running a new local server, and should sandbox servers with minimal default privileges.
- Log everything an agent does, with the identity that did it.
Network design for agent containment
Network design decides what a fooled agent can reach and where it can send things. Permissions limit what an agent is allowed to do; the network limits what it is able to do, which still holds when a permission was set too broadly.
Keep agents that work with operational systems, such as warehouse handhelds, field devices, kiosks, vehicles and machines, on a local network where they coordinate with the devices around them. Give that network one narrow gateway to the internet, carrying only the messages the work needs. An agent that is manipulated by something it read cannot reach systems that are not on its network and cannot send your data to an outside address it has no route to.
Safe agent coordination with Offline Protocol
Offline Protocol is the coordination layer for that design. Agents and apps on phones, computers and gateways advertise what they can do, discover each other up to 8 hops away and call each other request and response over Bluetooth LE, Wi-Fi and local IP, with no broker, no registry server and no internet. Every agent’s device carries an Ed25519 identity it generated itself, verified device to device, and request and response bodies are sealed in an MLS session once one is confirmed. The provider decides which requesters it serves, and the requester keeps a list of approved providers, so authorisation stays in your hands.
When the site needs the wider network, it reaches it through a gateway provisioned against a defined contract, not through whichever device happens to have a signal. Your agents get the speed of coordinating with the machines next to them, and your company keeps a short, known list of doors.