HP Study Finds AI “Vibe-Hacking” Attacks Slipping Past Enterprise Defences

HP has released its latest Wolf Security Threat Insights Report, covering the October to December 2025 quarter, with findings showing that cybercriminals are increasingly leaning on AI tools and cheap, off-the-shelf malware components to build attacks faster and at lower cost – and that these low-effort campaigns are still managing to bypass enterprise security tools.

The report, based on data gathered from millions of endpoints running HP Wolf Security, identifies three notable attack campaigns from the quarter: a modular “flat-pack” malware approach using reusable components, a PDF-based infection chain that exploits a redirect to Booking.com to disguise malware delivery, and a fake Microsoft Teams website that installs the OysterLoader backdoor alongside the legitimate application.

Alex Holland, Principal Threat Researcher at HP Security Lab, put it in straightforward terms. “It’s the classic project management triangle – speed, quality and cost. You often sacrifice one of them. What we’re seeing is many attackers are optimising for speed and cost, not quality. They are not using AI to raise the bar; they’re using it to move faster and reduce effort. The campaigns themselves are basic but the uncomfortable reality is they still work.”

Flat-Pack Malware Reduces the Effort to Build Campaigns

One of the report’s central findings is the rise of what HP researchers describe as “flat-pack” malware: attack campaigns assembled from modular, inexpensive components purchased from hacking forums, requiring only minor adjustments to be operational.

In Q4 2025, HP’s researchers observed multiple campaigns that used different initial infection file types, social engineering methods and final payloads, yet all relied on identical intermediate stages to install malware on endpoints. A recurring element involved downloading images from archive.org and using them to conceal malicious code.

Attackers emailed archives containing obfuscated malicious scripts and SVG files to infect PCs. The scripts were disguised as scanned Word documents using a double file extension trick, encouraging recipients to open them. Once executed, the scripts downloaded images from archive.org that contained Base64-encoded .NET malware hidden inside, extracted using reflective loading techniques.

Hosting malicious content on archive.org offers attackers several advantages: it removes the need to maintain their own infrastructure, avoids exposing payment or hosting records that could identify them, and increases the chance the download will succeed, since trusted domains are less likely to be blocked by web traffic filters.

The final payload in some of these campaigns was DarkCloud, an information stealer with Telegram-based remote access capabilities. Operators issue commands and monitor infected PCs through a Telegram channel, while the malware collects browser credentials and email data. In campaigns targeting organisations in Latin America, attackers used AsyncRAT instead, with lures shifting from government impersonation in earlier quarters to mimicking local banks by Q4.

Vibe-Hacking Scripts Exploit Booking.com Redirects

The report also documents a technique HP researchers are calling “vibe-hacking,” where attackers use AI tools to generate ready-made infection scripts – a nod to the “vibe coding” trend in legitimate software development, where developers rely on AI to write code without deep technical knowledge.

In one campaign, a fake invoice PDF was used as the lure. A link inside the PDF directed victims to a compromised website that silently triggered a malicious download before immediately redirecting the browser to Booking.com, creating the impression that the trusted platform had initiated the file download.

The attackers also used double file extensions with space padding to make the downloaded file appear to be a PDF rather than an executable script. The infection chain involved JavaScript that downloaded and executed a PowerShell stage – and unlike the obfuscated JavaScript, the PowerShell code contained no obfuscation at all, and was extensively commented in a way that suggested it was either purchased as a customisable toolkit or generated with the help of a generative AI tool.

“We are seeing more malicious PowerShell code in the wild that is over-commented in this way,” the report states, pointing to template-style scripts and verbose comments as hallmarks of AI-assisted development being adopted in the cybercrime space.

The final payloads in these campaigns were Formbook and XWorm, both capable of stealing browser credentials, cookies and system information, and providing remote access to infected systems.

OysterLoader Hidden Inside Fake Microsoft Teams Installer

The third major campaign documented in the report used search engine poisoning and malicious advertising to drive victims to a fake Microsoft Teams download site. The site closely mimicked the legitimate Microsoft Teams page, with only the URL revealing its true nature.

When a victim clicked to download the installer, the site generated a download link in the background by sending an HTTP OPTIONS request to verify that the payload domain was still active. The installer itself contained the legitimate Microsoft Teams application alongside several additional hidden components.

Among those components was an executable named “dwr.exe,” a legitimate, signed component from the video editing application CapCut. The installer also bundled a modified DLL named “mpr.dll” carrying the malware. When the installer ran, the CapCut executable loaded the modified DLL through a technique known as DLL sideloading, which allows malicious code to execute within a trusted, signed process, helping it bypass Windows security checks.

The malware installed was OysterLoader, a backdoor that gives attackers control over the device and allows additional malware to be deployed. HP’s researchers noted that OysterLoader is frequently seen in the lead-up to ransomware deployment.

Because the real Microsoft Teams application was installed as expected, users were less likely to notice anything suspicious had occurred.

Office Macros Persist in Asia Pacific

While macro-based attacks have declined globally, HP’s research found they remain active in the Asia Pacific region. Two campaigns in the dataset used different initial file types – Word documents and Excel spreadsheets – but led to the same infection chain, with each document using social engineering images to persuade recipients to enable Visual Basic for Applications (VBA) macros.

The macros downloaded PowerShell scripts that closely matched the structure seen in the Formbook and XWorm campaigns, with the key difference being the use of an RC4 decryption algorithm rather than XOR. The final payload in both the Word and Excel campaigns was Agent Tesla.

Agent Tesla was configured to harvest email contacts from compromised devices and send them to operators via a Telegram command and control channel. Keylogging and screen capture features were disabled in this configuration, suggesting the campaign’s primary goal was building a list of new targets for follow-on attacks.

Detection Rates and Threat Vectors

The report’s data from Q4 2025 shows that at least 14 per cent of email threats identified by HP Sure Click bypassed one or more email gateway scanners, a rise of 3 percentage points compared to Q3.

Scripts and executable files were the most common malware delivery type, accounting for 38 per cent of threats caught by HP Sure Click, up 8 percentage points on Q3. Archives were the second most common at 36 per cent, down 9 percentage points. The top archive formats abused by attackers were ZIP, RAR, GZ, 7Z and TAR. Documents such as Word formats accounted for 11 per cent of threats, malicious spreadsheets 4 per cent and PDF files 8 per cent.

Email remained the leading delivery vector, responsible for 58 per cent of threats, though that figure fell 9 percentage points compared to Q3. Malicious web browser downloads accounted for 23 per cent of threats, growing 7 percentage points. Threats via other vectors, such as removable media, made up the remaining 19 per cent.

HP Wolf Security customers have collectively clicked on more than 60 billion email attachments, web pages and downloaded files with no reported breaches, according to the company, owing to the platform’s approach of isolating threats inside secure micro-virtual machines rather than relying solely on detection.

The Limits of Detection-Based Defences

Dr. Ian Pratt, Global Head of Security for Personal Systems at HP Inc., argued that the findings point to a fundamental limitation in how most organisations approach endpoint security. “AI-assisted attacks are shining a spotlight on the limitations of detection-led security. When attackers can generate and repackage malware in minutes, detection-based defences can’t keep up.”

Pratt called for a shift in approach. “Instead of trying to spot every variant, organisations need to reduce exposure. By containing high-risk activities – like opening untrusted attachments or clicking unknown links – within an isolated environment, businesses can stop threats before they cause damage and remove an entire class of risk.”

Last Updated on March 6, 2026 by Nick Ross