Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

AI agent left SOC 27 minutes to respond versus 53 for a junior attacker

A SOC is the team that watches the alerts coming from security systems. The author reproduced an attack on a test network: an AI agent running Claude Sonnet 4.6 and a junior pentester both hunted for a flag on the domain controller. The alert fired for both after ~3 minutes, but after that the human attacker had ~53 minutes left on the clock, while the agent had ~27. If your response from alert to block takes longer than half an hour, only automation can stop an attack like this.

AI-processed from Habr AI; edited by Hamidun News
AI agent left SOC 27 minutes to respond versus 53 for a junior attacker
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Alerts~300 for the agent versus ~100 for the human, but the noise doesn't slow it down
Honeypot accountany access is grounds to isolate the host immediately
Key metric #3Two signals from one account within a few minutes: automatic block

Picture a night like this: the security system raises an alert at 03:12, the on-duty tier-1 analyst opens it at 03:30, sees a suspicious request to a web portal and goes to look at neighboring events. By then the attack is already over. Nights like this used to end fine: a human attacker has to look around, pick a tool, make mistakes and redo things, and those hours gave the SOC (security operations center) team enough time for the cycle of "spot it, check it, escalate it, block it."

An AI agent covers the same path without pauses, and that buffer is gone. The author of a write-up on Habr measured this on their own test bench: after the first alert, defenders had about 53 minutes against a novice human attacker and about 27 against an agent on Claude Sonnet 4.6.

If you are responsible for security, or simply build infrastructure, this analysis will help you decide which responses to hand over to automation.

The metric nobody usually counts

SOCs are usually measured by MTTD and MTTR: mean time to detect and mean time to respond to an incident. The author proposes looking at something else, the "response window." It is the time between the first alert in the SIEM (the system where events from all nodes are collected) and the moment the attacker reaches the goal: steals data, becomes a domain administrator or launches ransomware. An MTTR of 40 minutes is a good number as long as the window is measured in hours. If the window shrinks to half an hour, the same MTTR means you are not fast enough: the SOC is no longer stopping the attack but cleaning up after it.

Attackers sped up even before agents. According to CrowdStrike, the average time from initial access to moving to neighboring machines is 29 minutes (48 in the previous report), with a minimum of 27 seconds. According to Mandiant, an access broker hands access to the next group in 22 seconds, versus more than 8 hours in 2022, and vulnerabilities are exploited 7 days before a patch is released. Unit 42 records the fastest incidents at 72 minutes from entry to data theft, and CrowdStrike notes cases where theft began 4 minutes after entry.

Confirmed attacks involving agents are still few. In November 2025, Anthropic described a campaign called GTG-1002: the AI performed 80–90% of the tactical work, a human made 4–6 key decisions, there were about 30 targets, and at peak the agent sent several requests per second. The model vendor noticed the campaign, not the victims' SOCs.

Sysdig analyzed an AWS breach in which less than 10 minutes passed between leaked keys and administrator privileges, with 19 principals compromised. At a Russian company, according to PT ESC, about 3 hours passed between entry and data theft, and the attack was noticed because of its abnormal speed. Mandiant, however, is candid: 2025 has not yet become the year of AI-driven breaches, and most attacks still rely on human and configuration errors.

What the test bench showed

The author set up a typical small-company network: an OPNsense firewall with Suricata IPS, a web portal behind ModSecurity, Active Directory on Windows Server 2025, a server with PostgreSQL and Gitea, and a Windows 11 workstation. Defenses were Microsoft Defender with ASR rules, Sysmon and auditd, with events collected by the Wazuh 4.14 SIEM. Both attackers got the same one-sentence task: find the flag on the domain controller. Neither knew the environment. The agent ran on Claude Sonnet 4.6 (not a top-tier model), and the human was at junior or junior+ level.

The first alert arrived after about 3 minutes for both. The junior reached the domain controller in ~56 minutes, giving a response window of about 53. The agent did it in ~30 minutes, with a window of about 27. The gap is nearly twofold, which is not the chasm vendors like to paint. The agent wins not through the speed of any single step but through the absence of pauses: it doesn't read documentation and doesn't get tired at three in the morning. Each attacker had a single run, so this is an illustration, not statistics.

Their behavior differed. The agent went straight for the goal, hit rate limits and simply kept going, wrote scripts on the fly and probed even closed ports. The junior thought about evading defenses and reached for ready-made open-source tools. As a result, the agent triggered about 300 alerts and the human about 100. The agent is noisier, but the noise doesn't hinder it.

How to defend

The author suggests catching not individual actions but rhythm: a single agent command looks like a human one, but a queue of them within seconds gives the machine away. The rules are given in Sigma format; the thresholds are illustrative.

  • Five or more reconnaissance commands (whoami, net, nltest, klist) from one account on one host within a minute. Logon scripts and inventory tools will produce false positives.
  • pip install and apt install on servers: the agent installs missing libraries on the fly.
  • A honeypot account such as svc_backup_old with an SPN, a long password and no privileges. No one legitimately touches it, so any logon or Kerberos ticket request is grounds to isolate the host or disable the account immediately.
  • Send one signal to a human; send two signals from one account within a few minutes to automated response.
  • Do not automatically isolate domain controllers and critical servers: that would take down the whole network, and in some events the source address can be spoofed to point at your DC. For these, only raise an alert.
  • The "error and fix within seconds" pattern: two runs of nearly the same command a few seconds apart. Sigma can't express it, but it is visible by eye during incident review.
  • Enable the archive of all events in Wazuh (logall_json): by default only alerts reach the index, and without the archive you can't reconstruct the rhythm of an attack.

Limitations and takeaway

The experiment had one run, one agent and one junior, and the Sigma rules are marked experimental. The practical takeaway is simple: if half an hour is enough for an attacker and your response takes 40 minutes, automation has to take the first step, and humans should be left to decide the disputed cases. The response window had already shrunk before AI; agents simply removed the pauses from the attack.

Story

Key Questions Answered

What are the key details on The metric nobody usually counts?

SOCs are usually measured by MTTD and MTTR: mean time to detect and mean time to respond to an incident. The author proposes looking at something else, the "response window." It is the time between the first alert in the SIEM (the system where events from all nodes are collected) and the moment the attacker reaches the goal: steals data, becomes a domain administrator or launches ransomware. An MTTR of 40 minutes is a good number as long as the window is measured in hours.

What the test bench showed?

The author set up a typical small-company network: an OPNsense firewall with Suricata IPS, a web portal behind ModSecurity, Active Directory on Windows Server 2025, a server with PostgreSQL and Gitea, and a Windows 11 workstation. Defenses were Microsoft Defender with ASR rules, Sysmon and auditd, with events collected by the Wazuh 4.14 SIEM. Both attackers got the same one-sentence task: find the flag on the domain controller. Neither knew the environment.

How to defend?

The author suggests catching not individual actions but rhythm: a single agent command looks like a human one, but a queue of them within seconds gives the machine away. The rules are given in Sigma format; the thresholds are illustrative. Five or more reconnaissance commands (whoami, net, nltest, klist) from one account on one host within a minute. Logon scripts and inventory tools will produce false positives. pip install and apt install on servers: the agent installs missing libraries on the fly.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…