Stage 3 of 3 · Act
Which defenses stop AI-driven attacks, and which to fund first
Attackers now use AI agents to find a way in and spread within minutes. Here are the defenses you can put in place, from basic controls to AI tools of your own, and a simple way to decide which fit your company.
- 9 min read
- For CEOs, boards and CIOs
- Technical parts marked
Board questions this part answers
- What can each of our AI agents actually do, and can we stop it?
- Which risks have we accepted, and who owns them?
The 6 jobs of cyber security, and the frameworks for each
The most widely used map is NIST's Cybersecurity Framework 2.0. It sorts security into 6 jobs, and the other frameworks your team or auditor mentions fit inside them. Tap a job to see what it asks of leaders and which frameworks go deeper on it. 61
Tap a job on the ring, then switch on the AI layer
Govern sits at the center and steers the other 5.
NIST's draft Cyber AI Profile adds an AI lens to each job: secure your own AI, use AI to defend, and prepare for attacks that use AI. 62
Govern
Decide who owns cyber risk and how much of it you accept.
What it asks of leaders
Name an executive owner, set a risk appetite the board signs, and review it every quarter. This is the job added in 2024, and it puts cyber risk on the board's agenda.
With AI in the picture
Decide which AI tools and agents the company may use, who approves each one, and what each may do.
Frameworks that go deeper
Identify
Know what you have and what matters most.
What it asks of leaders
Keep a list of critical services, systems, suppliers and AI agents, each with an owner, and add to it whenever something new goes live.
With AI in the picture
Add every AI model, assistant and agent to the list, with the data and tools it can reach.
Frameworks that go deeper
Protect
Make the common attacks fail.
What it asks of leaders
Strong sign-in, fast patching, limited admin rights and payment rules that one person cannot override.
With AI in the picture
Give agents their own identities and the smallest set of permissions their job needs.
Frameworks that go deeper
Detect
Notice an attack in minutes.
What it asks of leaders
Someone, in-house or a provider, watches alerts day and night and has authority to act.
With AI in the picture
Use AI to sort alerts, and watch your own agents for actions outside their job.
Frameworks that go deeper
Respond
Contain it and keep the business running.
What it asks of leaders
A rehearsed plan: who leads, who decides, who tells regulators and customers, and who may switch systems off.
With AI in the picture
Decide in advance which containment steps AI may take alone and which need a person.
Frameworks that go deeper
Recover
Get back to normal, and prove it.
What it asks of leaders
Backups kept offline and tested by restoring them, plus a manual way to run the most important services for a week.
With AI in the picture
Keep a known-good copy of each agent's instructions and permissions, so you can roll it back.
Frameworks that go deeper
Go deeperWhich framework is for what
| Framework | The question it answers | Who uses it | When you need it |
|---|---|---|---|
| NIST CSF 2.0 | Are we covering all six jobs? | Boards and security leaders | As the shared map for everyone |
| CIS Controls v8.1 | What exactly should we do first? | IT teams | To turn the map into a task list, in 3 levels |
| ISO/IEC 27001 | Can we prove it to others? | Companies selling to large clients or abroad | When customers or regulators ask for certification |
| Zero Trust | What if someone is already inside? | IT architects | When moving to cloud or opening systems to partners and agents |
| OWASP Top 10 for agentic apps | How do AI agents get abused? | Teams building or buying agents | Before any agent can act on its own |
| MITRE ATT&CK and ATLAS | How do attackers actually work? | Security teams and providers | To test detection against real attacker behavior |
| ISO/IEC 42001 | Do we manage AI responsibly? | Companies with many AI uses | When AI becomes part of core operations |
| UU PDP, OJK and BI rules | What does the law require of us? | Legal, compliance and IT | Always, in Indonesia |
5 levels of defense, and the order to fund them
Each level builds on the one below. Most companies should finish level 1 before buying anything at level 3. Tap a level to see what it stops and what it takes.
Tap a step to see what each level stops
Level 1
Essentials
Close the doors most attacks use
What it stops
Stolen passwords, fake sign-in pages, known flaws, payment fraud by email, and ransomware that would otherwise wipe out your only copy.
- Multi-factor sign-in on email, admin and remote access. In a large study of accounts showing suspicious activity, it cut the risk of takeover by 99.22%. 69
- Critical flaws fixed within days, and a list of every system that faces the internet.
- Offline backups, tested by restoring them.
- A callback and a second approver before any change to bank details or any large payment.
What it takes
One capable IT lead and a few weeks. Mostly settings and habits, with little new to buy.
Before you start
Nothing. Every company starts here.
UK companies certified to a similar set of basic controls make 92% fewer cyber insurance claims. 70
Level 2
Managed protection
Someone watching, day and night
What it stops
Attacks that get past level 1 and would otherwise run unseen for weeks.
- Endpoint detection and response (EDR) on every laptop and server, to spot and stop suspicious behavior.
- Email security that checks links, attachments and sender identity.
- A managed detection and response (MDR) provider whose team watches your alerts 24/7 and can isolate a device at 3 a.m.
What it takes
A provider contract and an internal owner who answers their call.
Before you start
Level 1 in place, so a stolen password alone is not enough to sign in.
In 2025, 52% of intrusions were found by the victim's own team. The rest were reported by someone outside. 73
Level 3
AI-assisted defense
AI sorts the noise so people act faster
What it stops
Fast, high-volume attacks that swamp human analysts, and convincing AI-written phishing.
- AI triage, where AI reads each alert, gathers the evidence and proposes a verdict for an analyst.
- AI email screening that judges the writing, the request and the sender's normal behavior.
- Identity monitoring that flags a sign-in or permission change unusual for that person.
What it takes
Clean data from level 2 tools, and analysts who check the AI's work. Ask your MDR or EDR provider whether it is included.
Before you start
Level 2 tools feeding the AI. The AI can only judge the alerts and logs those tools send it.
Companies using security AI and automation saved US$1.93m per breach on average in 2026. 4
Level 4
Autonomous response
Contain an attack in seconds
What it stops
Attacks that spread faster than a person can wake up and log in.
- Pre-approved actions the system takes alone: isolate a laptop, suspend an account, block a sign-in.
- AI agents that investigate an incident end to end and hand a person a summary and a recommendation.
- Automated attack testing, run every month, to prove which gaps an attacker could really use.
What it takes
Written rules for what AI may do alone, tested rollbacks, and a security team or provider able to oversee it.
Before you start
Level 3 running well, and a business owner who has signed off on the automatic actions.
Criminals took an average of 29 minutes to move from the first machine to others in 2025. 7
Level 5 · Alongside any level
Securing your own AI
Keep your agents inside their limits
What it stops
Hidden instructions in documents, agents with too much access, and AI tools staff use without approval.
- One register of every AI agent, with an owner, its data, its tools and a tested way to stop it.
- Guardrails and gateways that check what goes into and comes out of each agent, and block risky actions.
- A separate identity for each agent, with the least access its job needs, and a person to approve payments, deletions and data leaving the company.
What it takes
Applies at any size, as soon as an AI agent can act on its own. Effort grows with the number of agents.
Before you start
A decision on which agents may act at all. See the authority ladder in the AI agents chapter.
Of organizations that had an AI-related breach in 2026, 9 in 10 lacked proper access controls on their AI. 3
8 kinds of AI that defend you
Each card is a type of tool, and most security vendors now sell several types. Tap one to see what it does, the evidence so far, and the question to ask before you buy.
Tap a kind of AI to open it
AI email and payment-fraud screening
Businesses reported US$3.05bn lost to email fraud to the FBI in 2025. 5
What it does
Learns how your suppliers and executives normally write and what they normally ask for. A request to change bank details from a lookalike address, or in an unusual tone, is held for checking.
Where it falls short, and what to ask
It cannot stop a fraud that arrives by phone or WhatsApp. Keep the callback rule. Ask how many of last month's held emails were real fraud, and how many were false alarms.
Useful from Level 3, for any company that pays suppliers by transfer
Deepfake voice and video checks
A finance employee in Hong Kong paid out about US$25m after a video call with deepfaked colleagues. 24,25
What it does
Analyzes a call or recording for signs of synthesis and warns the participant. Some tools add a verified badge to genuine internal callers.
Where it falls short, and what to ask
In a test on deepfakes circulating in 2024, open-source detectors lost about half their accuracy. Ask how the tool was tested on deepfakes from the last six months. Keep a code word or callback for large requests. 72
Useful from Level 3, for finance teams and executives
AI security analysts
In a randomized trial with 167 security analysts, those working with an AI agent found up to 6.5× as many real threats per minute. 71
What it does
Each alert gets an investigation: what happened, on which machine, whether it matches known attacks, and what to do. A person approves the verdict.
Where it falls short, and what to ask
Works only as well as the alerts and logs it can see. Ask what happens when the AI is unsure, and who checks the cases it closes.
Useful from Level 3, often inside an MDR service
Autonomous containment
What it does
When an attack is confirmed, the system isolates the laptop, suspends the account or blocks the sign-in by itself, then tells a person.
Where it falls short, and what to ask
A wrong decision can stop a factory line or lock out your CEO. Ask which actions run alone, which need approval, and how fast each can be undone.
Useful from Level 4
Identity threat detection
Criminals who talked IT help desks into resetting passwords were behind the 2023 attack on MGM Resorts. 14,20
What it does
Watches sign-ins, password resets and permission changes for people, service accounts and AI agents. A help-desk reset followed by a new admin role at midnight is flagged.
Where it falls short, and what to ask
Needs every identity system connected, including old ones. Ask whether it covers service accounts and AI agents as well as people.
Useful from Level 3
Exposure management and AI testing
Exploited flaws were the top way in, at 32% of the intrusions Mandiant investigated in 2025. 73
What it does
Continuously maps what faces the internet, ranks flaws by how attackable they are, and runs safe simulated attacks to prove which ones matter.
Where it falls short, and what to ask
A list of flaws helps only when someone fixes them. Ask how findings reach the person who must fix them, and by when.
Useful from Level 2 for mapping, level 4 for automated attack testing
AI agent guardrails and gateways
OWASP's 2026 list puts agent goal hijack, where hidden text redirects an agent, at number 1. 65
What it does
Sits between agents and the tools and data they use. Screens incoming documents for hidden instructions, blocks actions outside the agent's job, and logs every step.
Where it falls short, and what to ask
No guardrail yet stops every hidden instruction. Ask which of the OWASP top 10 risks for agents it is tested against, and what it logs. 65
Useful from Level 5, as soon as an agent can act
AI data-loss prevention
Only 40% of organizations in IBM's 2026 study used access controls on their AI models and data. 3
What it does
Recognizes customer records, contracts or source code being pasted into an AI tool or sent out by an agent, and blocks or masks it.
Where it falls short, and what to ask
Staff can still use AI tools on personal phones. Ask whether it sees AI tools used in a browser, and whether it explains each block to the user.
Useful from Level 3, sooner if you handle personal data under UU PDP
Which attack responses AI may take alone, and which need a person
Let AI act alone only where the action is quick to undo. Agree these lines before you switch anything on, and review them every quarter.
- 7
29 min
the average time criminals took to spread to other machines in 2025.
- 73
14 days
the median time an attacker stayed unnoticed in 2025.
| Action | AI alone | AI proposes, a person approves | People only |
|---|---|---|---|
| Sort and summarize alerts | Yes | ||
| Block a known malicious email or website | Yes | ||
| Isolate one laptop | Yes, if it can be undone in minutes | ||
| Suspend an employee's account | Yes | ||
| Shut down a server or production line | Yes | ||
| Cut off a supplier's connection | Yes, out of hours | Yes, in hours | |
| Tell customers, regulators or the police | Yes |
The same thinking applies to your business AI agents. The authority ladder in the AI agents chapter sets how much each one may do.
6 questions to find your level, and who should run it
Answer 6 questions about your company. You get the level to aim for this year, a way to resource it, and 3 first moves. Nothing you enter leaves this page.
Answer 6 quick questions about your company
Go deeperHow the result is worked out
Company size and IT team each score 0 to 3. Level 2 applies when either scores 1 or more, when you are regulated, or when the business stops within hours. Level 3 applies when the two together reach 3, or when a regulated company has 50 people or more. Level 4 applies when you have dedicated security staff and either 250 people or more or a regulator. Level 5 is added whenever agents take actions. The way to resource it follows your IT team. The moves start at level 1 unless all three basics are in place, then follow the levels upward. When your agents take actions, securing them becomes the second move. Treat it as a starting point for a conversation with your CIO. It is not professional advice.
7 questions to ask before you buy any security tool
Every vendor in these categories will say its product uses AI. These questions show what it actually does, and what you get if you leave.

