Back Linkedin

An AI-Assisted Cyber Attack: Inside a Unit 42 Investigation | Wendi Whitmore

Linkedin • September 2, 2026

investigated a break-in this month that was over in less than ten hours. The same work would have taken a skilled human team roughly two weeks.

A single attacker found an opening in a public-facing application and got inside. From there, they did not do any manual work themselves. They pointed AI agents at the problem and let them run.

The agents systematically mapped the company's internal systems. They read through the company's own source code, hunting for passwords and keys developers had left behind, and they found them. Those keys opened the vault where the company stored its most sensitive credentials, and from that point forward the attacker had free rein.

Then came a new twist. The agents took the keys to the company's own AI systems and used them to power the rest of the attack. The victim's technology ran the intrusion, and the victim paid the compute bill for the privilege.

The attacker's parting move was to instruct the AI to generate an 80-page report on everything wrong with the company's security and leave it behind.

One control did hold. When the agents tried to bury a backdoor in the company's infrastructure code, a rule requiring human review blocked it. Good controls still work. What has to change is that they can no longer depend on someone noticing in time.

For executive leaders reading this: the response window we used to plan against was measured in days. This incident argues that window is now measured in hours. That single shift changes what "prepared" actually looks like across most enterprise environments.

Three questions worth putting to your security leader this quarter.

Do we know where every AI system, key, and connection in our environment actually lives, and who has access to use them?

If we found an intruder in our environment right now, how do we contain them & how do we prevent their access to our AI systems?

Are we running AI to hunt through our code for exposed secrets, embedded credentials, and vulnerabilities, before attackers point their agents at the same targets?

The full story

This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.

AI-Driven Ransomware Attack Completes in Under 10 Hours On September 2, 2026, a human attacker utilized frontier AI agents to execute a ransomware attack on an enterprise, completing the breach in less than 10 hours. The attack involved over 50 techniques from the MITRE ATT&CK framework, which would typically require two weeks of manual effort. Initial access was gained through a public API endpoint, followed by automated reconnaissance and credential harvesting from source code repositories. The AI agents gained root access and hijacked CI/CD pipelines, using the victim's own AI infrastructure to further the attack. The attacker left behind an 80-page security audit detailing the vulnerabilities exploited. This incident marks a significant escalation in the use of AI in cyberattacks, demonstrating operational efficiency without the need for novel exploits. The attack highlights the urgent need for organizations to adapt their security measures to counteract AI-assisted threats. 23 sources 71

Extracted Entities

CWE Weaknesses (2)

CWE-200 - Exposure of Sensitive Information CWE-798 - Use of Hard-coded Credentials

MITRE ATT&CK (1)

T1552.001 - Credentials In Files