AI is now doing the hacking. Defence has to move at machine speed too
Syria and the region
Mandiant, Google's threat intelligence team, CrowdStrike, Trend Micro and Trellix all report the same shift: attackers now hand large parts of a break-in to AI agents that work in seconds, not days. Here is what they found, what it means for Syria and the region, and how to defend at the same speed.
CyberSyria Threat Intelligence · Published 11 November 2026
- minutes: the fastest jump from one stolen account to deep access (Mandiant)
- 15minutes: the fastest jump from one stolen account to deep access (Mandiant)
- attacker actions in just 23 seconds, run by one AI agent (Mandiant)
- 27attacker actions in just 23 seconds, run by one AI agent (Mandiant)
- AI-agent intrusions in the second half of 2026 compared with the first (CrowdStrike)
- ×3AI-agent intrusions in the second half of 2026 compared with the first (CrowdStrike)
- AI agents working together in one attack toolkit (Trend Micro)
- 44AI agents working together in one attack toolkit (Trend Micro)
Key takeaways
- 1
Attackers now let AI agents carry out the break-in: scanning, moving through the network and preparing data to steal, in minutes or hours instead of weeks.
- 2
People alone cannot keep up with attacks that move in seconds. Defenders need AI and automation that spot unusual activity and cut access automatically.
- 3
AI agents still walk through old doors: unpatched systems, weak or reused passwords, and admin pages open to the internet. Closing those doors is the cheapest defence of all.
What happened in recent weeks
Mandiant: AI "pentest agents" are being used in real break-ins
What happened
Mandiant and Google Threat Intelligence Group report that since late August 2026 attackers have used open-source AI agent tools, built for security testing, in real intrusions. Tools named include ARTEX, PentAGI, Hexstrike and CAIRN. The agents handled moving through networks, breaking into databases and preparing data for theft. One group went from a single compromised account to deep access in about four hours, and once in 15 minutes. The agents act in bursts of a few seconds (27 actions in 23 seconds in one case) and fix their own mistakes. They are noisy and leave many logs, but they work. In one case an agent used a company's own notification system to email more than 1,300 staff. Users include extortion gangs and suspected China-linked espionage groups. Google assesses with high confidence that criminals will keep adopting these tools.
Why it matters for Syrians
These tools are free and public, so they lower the skill needed to attack. Syrian organisations that are rebuilding their systems, often with small IT teams and no night shift, are exactly the kind of target where a fast, noisy attack can finish before anyone looks at the logs.
What to do
- Keep admin consoles, cloud dashboards and build systems off the open internet; reach them through a VPN.
- Replace long-lived passwords and API keys with short-lived ones that only work for one job.
- Put limits on internal systems that can send mass emails or export data.
Google: "assume that all threat actors are using AI"
What happened
Google Threat Intelligence Group's quarterly AI Threat Tracker describes a move "from prompting to autonomy". State-backed groups linked to Russia, China, Iran and North Korea use AI models for research, fake personas, writing attack scripts and troubleshooting intrusions. In one case an attacker used an AI agent framework to plan and run a mass password-harvesting operation in under six hours. Attackers also target AI tools themselves: poisoned software packages and plugins aimed at AI coding assistants, and stolen AI accounts and cloud capacity. Google's John Hultquist summed it up: "At this point, we can assume that all threat actors are using AI in some capacity."
Why it matters for Syrians
Fake personas and fluent messages in any language make phishing in Arabic far more convincing. And many Syrian developers and students now use AI coding assistants and free plugins, which are a new way in.
What to do
- Treat a perfect, friendly message that asks for a code, a payment or a file with more suspicion, not less.
- Install AI plugins, "skills" and packages only from official sources, and check what they are allowed to do.
- Protect AI and cloud accounts with two-step verification and never paste API keys into code or chats.
CrowdStrike: four new groups let AI agents run their attacks
What happened
CrowdStrike found four previously unreported groups using AI coding assistants to make decisions and run commands during intrusions with little human input. They tricked the AI's safety rules by pretending to run authorised security tests. In one attack on a hospital in Argentina, an agent switched route when blocked, reached the network's main login server and stole password data in about 20 minutes. One group hit at least 15 healthcare organisations in Argentina. CrowdStrike's hunting team says AI-agent intrusions in the second half of 2026 are already three times the first half, and expects them to keep growing.
Why it matters for Syrians
Hospitals and clinics are a favourite target because they cannot afford downtime. Syrian health providers are digitising fast, often on old systems.
What to do
- Set rules that disable an account or isolate a device automatically when activity looks abnormal, without waiting for a person.
- Protect the main login server (domain controller) like the crown jewels: few admins, strong 2FA, close monitoring.
Trend Micro: a 44-agent AI toolkit hit at least 19 organisations
What happened
Trend Micro researchers studied three attacker servers that had been left exposed. In one, an open-source tool ran 44 AI agents covering the whole attack, from scanning to stealing data and writing a neat report on each victim. It also built fake Japanese-language login pages automatically. At least 19 organisations were targeted in October, mostly in Japan. In a separate campaign, where AI was used mainly to build the attacker's tools, the target list included organisations in Qatar, the UK, France, Italy and Colombia.
Why it matters for Syrians
If AI can build convincing fake login pages in Japanese, it can do the same in Arabic. And the Gulf is already on the target lists.
What to do
- Use passkeys or an authenticator app, which fake login pages cannot easily steal, instead of SMS codes.
- Patch internet-facing devices (VPNs, firewalls, web servers) first: they are what the scanners find.
Trellix: AI has become "an autonomous operator" in cybercrime
What happened
Trellix's November threat report says AI "has evolved from a drafting aid into an autonomous operator". Underground markets now sell ready-made attack agents, not just chatbots that write scam emails. One ransomware service claims to use AI to sort stolen files and set the ransom based on the victim's legal and financial exposure. Trellix notes these claims are not verified. Trellix expects nearly every capable attacker to use AI and calls for detection and response at machine speed.
Why it matters for Syrians
When attack agents can be bought, small criminal groups get capabilities that used to need skilled teams. Size is no longer protection.
What to do
- Keep offline backups and test restoring them, so a ransom demand is never your only option.
Trend to know
Fight machine speed with machine speed, on top of strong basics
When an attack runs 27 actions in 23 seconds, no analyst can read the alerts fast enough, and an approval that waits until morning comes too late. Defenders need AI on their side: tools that learn what normal looks like and flag a sudden burst at once, automatic rules that lock an account or isolate a laptop in seconds, and AI assistants that sort thousands of alerts so people can focus on the few that matter. Defenders can also use AI to find their own weaknesses first: Mandiant reports that its AI code review found more than 100 critical flaws in two days during one investigation. But AI defence is not magic. The same reports show that agents still get in through unpatched systems, weak passwords and exposed admin pages, and that they are noisy. Strong basics make them easy to stop, and AI makes stopping them fast.
How to defend
For everyone
- Turn on two-step verification everywhere, ideally with passkeys or an authenticator app rather than SMS.
- Use a different password for every account, kept in a password manager.
- Update your phone, computer and browser as soon as updates arrive.
- A perfectly written message is no longer a sign it is real. Check urgent requests through another channel.
For organisations
- Detect at machine speed: use security tools (EDR/XDR) that learn normal behaviour and flag sudden bursts, such as dozens of commands in seconds or one account logging in everywhere.
- Respond automatically: let those tools isolate a device or disable an account in seconds, then let a person review. A 15-minute attack does not wait for approval.
- Set traps: plant fake "canary" passwords, keys and files. AI agents grab everything they find, so the moment one is used you get an alarm.
- Find your holes first: use AI-assisted code review and vulnerability scanning on your own systems before attackers do.
- Shrink the blast radius: short-lived keys, least privilege, network segments, and rate limits on systems that can send mass emails or export data.
- Close the old doors: patch internet-facing systems first and take admin pages off the open internet.
What to watch in the coming months
- AI-written phishing and fake login pages in Arabic aimed at Syrian banks, aid groups and government services.
- Fake AI tools, plugins and "skills" that hide malware, aimed at developers and students.
- Theft of AI accounts and API keys to run attacks on someone else's bill.
How we made this report
This report uses public sources only: news outlets, official statements, national cyber authorities, internet measurement data and published research, in Arabic and English. We label how sure we are about each item, name who is behind an attack only when a credible published source does, and never link to leaked data. CyberSyria is non-political and vendor neutral. Classification: TLP:CLEAR, free to share.
Spotted a mistake or have a tip? Write to hello@cybersyria.org.
Get reports like this first
Join CyberSyria, the community for everyone in Syria and the diaspora who works in or cares about cyber security. Membership is free.
