Inside an AI Attack Against a US Retailer
When hand-typed commands failed to breach a US retailer, the attacker turned to AI-generated scripts, achieving lateral movement in minutes.
A threat actor deployed an AI model during a stalled cyberattack and rescued the campaign, achieving lateral movement in the process.
The attack, detailed today by Symantec, targeted an unnamed US retail organization in September. The threat actor’s hand-typed commands were repeatedly blocked, so the attacker switched to what appeared to be AI-generated commands and rapidly broke through the retailer’s defenses. The attacker ultimately harvested credentials and achieved lateral movement within minutes.
Symantec believes AI was used because the timing between commands was more indicative of automation than direct hands-on keyboard. That said, the timing was such that investigators believe this activity was “more consistent with an operator relaying output to an AI tool than with an autonomous agent.”
The security firm said this incident is one of dozens of AI-assisted attacks Symantec has investigated since May, though it added that activity was heavily concentrated around September.
A New Era for Attackers Leveraging LLMs
Threat actors have utilized large language models (LLMs) in their attacks as long as the technology has been widely available. Early applications involved generating convincing phishing lures or conducting basic data reconnaissance, but recent evidence shows attackers deploying AI to execute end-to-end attacks, assist with malware development, and more. More recently, attackers have abused — and continue to abuse — agentic penetration testing tools to break into target environments at machine speed.
In this attack specifically, Symantec said it did not identify the initial access vector, focusing more on the timeline it was able to glean from its post-incident investigation. According to the blog post, the attacker worked at first by hand, which investigators determined due to artifacts like typing errors.
When every staged payload was terminated by security software, indicators suggest attacker tactics changed when they tried again the next day. “For almost two hours,” Symantec said, “the attackers submitted long, error-free PowerShell scripts with labelled output sections and, in places, comments that read as a model’s reasoning.” Approximately 10 minutes after querying the victim’s SolarWinds Orion monitoring software, attackers managed to laterally move to a second server.
The blog post hypothesizes that the attackers behind the attack were likely financially motivated, as they searched point-of-sale scripts for passwords, and one address they contacted was linked to an operation that stole card data from online retailers.
Clutching Cyberattack Victory From Defeat
During the second-stage activity, the threat actor gained initial access to a Web server and stole credentials from configuration files tied to the SolarWinds Orion monitoring software. And while the idea of AI speeding up or improving attacker tradecraft is nothing new, the attack is its deployment of AI to a hands-on keyboard attack in progress and possibly seizing victory out of an apparent defeat. It presents LLMs as a difference-maker for a less skilled threat actor.
Dick O’Brien, principal intelligence analyst on the Symantec Threat Hunter Team, tells Dark Reading this incident is one of several the team has seen in recent weeks where AI revived an attack. While he calls this a game-changer for attackers, “defenders can potentially leverage that pattern” as AI behaves in a particular kind of way, such as testing variants of a technique in quick succession.
“The retry pattern is a signal itself,” he says. “Blocking on behavior rather than file names or exact commands matters much more. Treating a blocked attempt followed by a variant as a high-priority alert is worth considering.”
Although no vulnerability was exploited and no patching was required, O’Brien believes this trend of AI accelerating attacks is and will continue to affect the patching race. “Our view is that defenders need to start looking at mean time to mitigate rather than mean time to patch. Patching is still important, but it can’t be the only countermeasure.”
Originally published on Dark Reading
