AI GUIDEPartner content

Agentic Cyberattacks: How They Differ from Ordinary Automation

A look at a new threat landscape: an AI agent observes its environment, chooses the next step, and controls tools. What changes for defense.

Affiliate link: your price stays the same and the project earns a commission.

Ordinary automation performs predefined steps. An agentic loop can assess the result, choose the next tool, and restructure the sequence without a new human command. For cyber defense, this is a significant difference: speed increases not just at one stage, but across the entire chain of actions.

What Makes an Attack Agentic

A script is good at repeating a known procedure, but stops when it gets an unexpected response. An AI agent receives a goal, observes the state of the environment, reasons about the next step, launches a tool, and assesses the result again. If one path does not work, it can choose another.

This does not mean complete autonomy or infallibility. In practice, an agent may misinterpret a tool’s output, get stuck in a loop, or waste resources. However, even an imperfect system takes routine work off the operator’s hands: sorting results, repeating checks, preparing code, and managing a large number of parallel tasks.

What Incident Response Teams Have Already Observed

Google Threat Intelligence Group reported that malicious actors were moving from individual AI queries to multistage automation. In one case, a group was designing a dynamic agentic framework to test systems. In another, an attacker deployed a multi-agent pipeline in a compromised cloud environment to collect credentials at scale.

Google also described a dashboard that managed more than 23 800 discovered secrets, including keys for cloud and AI services. This figure refers to a specific server that was discovered, not an estimate of all such attacks. The provider reported blocking the related resources and updating its security mechanisms.

The main takeaway is not that AI invented a new class of vulnerabilities. It connected reconnaissance, testing, processing results, and managing findings into a faster loop.

Which Signals Are Changing for Defenders

In a manual attack, activity often comes in bursts: the operator studies the results and returns later. An agent can act steadily, quickly change parameters, and repeat attempts at scale. As a result, an individual event becomes less informative, while the sequence becomes more important.

It is worth correlating:

  • short bursts of similar requests to many hosts;
  • rapid switching between reconnaissance, access testing, and accessing secrets;
  • the creation of new processes or tasks immediately after an unusual network event;
  • the use of cloud addresses that are usually considered trusted;
  • access to token and key stores after a security rule is triggered.

Individually, these signs also occur during ordinary administration. Correlation by time, account, host, and target helps distinguish an engineer’s work from an automated attack.

Why Defense Must Break the Loop

Blocking one IP address does little good if the system can change its route. Banning one tool will not stop an agent that can find a substitute. A defensive measure is more effective when it deprives the loop of feedback: restricts outbound network traffic, makes a secret inaccessible, requires separate confirmation before changing the environment, or ends the session when the target changes.

It is useful to separate observation and action permissions. A service that analyzes logs does not need to be able to change configurations. An agent testing a test environment should not have access to production tokens. Even in the event of an error or compromise, this reduces the number of available next steps.

Five Measures That Make a Practical Difference

  1. Short-lived credentials. A token that is valid for a few minutes and only for one operation is less useful after it is stolen.
  2. Network segmentation. The test environment should not have an unrestricted route to production.
  3. Outbound connection controls. Permitted destinations are explicitly defined, and new addresses require verification.
  4. Event correlation. Identity, cloud, endpoint, and application logs are analyzed as a single chain.
  5. Budget-based stopping. Time, number of actions, volume of requests, and number of affected objects are limited.

These measures are useful regardless of whether an attack is controlled by a person, a script, or AI. Agentic scenarios simply make them more urgent.

How to Prepare for an Incident

The response plan should include not only blocking the account, but also revoking all related secrets, searching for similar activity patterns, and checking the cloud infrastructure from which the activity originated. If only the visible session is stopped, the autonomous pipeline may continue operating with keys it has already obtained.

Detailed observations and limitations of the source data are set out in the Google Threat Intelligence Group report. These are findings from Google and Mandiant investigations; they confirm the existence of specific campaigns, but do not provide comprehensive statistics for the entire internet.

Compare models before you start

The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.

Browse models

Affiliate link: your price stays the same and the project earns a commission.

agentic cyberattacks AI in cybersecurity autonomous agents defense against AI attacks

SEO Mind42 editorial team

We explore SEO and neural networks in practice: test services on our own projects, verify prices and limits against primary sources, and share things you can put to use the same day.

📚 Reference guide to SEO and AI 🔄 Materials are updated 🕐 Updated: 4 October 2026

Related reading

All in this section →