Fortinet spots new malware that corrupts its own headers to block forensic analysis, hide behavior, and communicate with its C2 server.

The FortiGuard Incident Response Team has released a detailed investigation into a newly discovered malware that managed to quietly operate on a compromised Windows machine for several weeks. What makes this malware different from others is its deliberate corruption of its own DOS and PE headers, a method designed to obstruct forensic analysis and reconstruction efforts by security researchers.

Despite this challenge, Fortinet’s team successfully obtained a memory dump of the live malware process, housed in a dllhost.exe process (PID 8200), along with a complete 33GB memory dump of the compromised system.

By carefully replicating the compromised environment, Fortinet’s researchers were able to bring the dumped malware back to life in a controlled setting, allowing them to observe its operations and communication patterns.

****Bringing Corrupted Malware Back Online****

Without its DOS and PE headers, the malware could not be simply loaded and executed like a normal Windows binary. The research team had to manually identify the malware’s entry point, allocate memory, and resolve API addresses that differed between the compromised system and the test environment. Through repeated debugging, address relocation, and parameter adjustments, they were finally able to emulate the malware’s behaviour in a lab setting.

The image shows the DOS and PE headers have been corrupted, which makes it challenging to fully reconstruct the executable from memory (Credit: FortiGuard)

According to Fortinet’s blog post shared with Hackread.com ahead of its publishing on Thursday, once operational, the malware revealed its communication with a command-and-control (C2) server at rushpaperscom over port 443, using TLS encryption.

Fortinet analysts traced the malware’s use of Windows API functions like SealMessage() and DecryptMessage() to handle encrypted traffic. They also identified an additional layer of custom encryption that wrapped specific data packets before applying TLS, further complicating traffic inspection.

****What the Malware Can Do****

Fortinet’s analysis confirms that the malware operates as a Remote Access Trojan (RAT), providing the attacker with several powerful features:

*   **Screen capture**: The malware takes periodic screenshots, compresses them as JPEGs, and sends them to the C2 server along with the titles of active windows.

*   **Remote server functionality**: The malware sets up a listening TCP port, allowing attackers to connect directly and issue commands or deploy additional attacks.

*   **System service control**: By interfacing with the Windows Service Control Manager, the malware can enumerate, manipulate, and potentially disrupt critical system services on the infected machine.

****How the Attack Works****

The initial infection relied on batch scripts and PowerShell to launch the malware, embedding it into a Windows process. Once running, the malware fetched the C2 server’s domain information from encrypted memory, established a secure connection, and began exfiltrating system details.

Full memory dump of the compromised machine. The image shows detailed file information for the “fullout” dump, used to recreate a local test environment for malware analysis. (Credit: FortiGuard)

During traffic analysis, Fortinet captured decrypted WebSocket requests and responses, uncovering how the malware collects and reports system information, including OS version and architecture.

Interestingly, the malware’s encryption scheme uses a randomly generated key for XOR-based scrambling of packet data before it is handed off for TLS encryption. This extra layer adds protection against simple network-based detection, forcing researchers to rely on endpoint inspection or memory-level analysis to catch malicious activity.

New Malware Spotted Corrupts Its Own Headers to Block Analysis

A recent investigation by cybersecurity researchers at Oasis Security has revealed a data overreach in how Microsoft’s OneDrive…

A recent investigation by cybersecurity researchers at Oasis Security has revealed a data overreach in how Microsoft’s OneDrive File Picker handles permissions, opening the door for hundreds of popular web applications, including ChatGPT, Slack, Trello, and ClickUp, to access far more user data than most people realize.

According to the report, the problem comes from how the OneDrive File Picker requests OAuth permissions. Instead of limiting access to just the files a user selects for upload or download, the system grants connected applications broad read or write permissions across the user’s entire OneDrive. This means that when you click to upload a single file, the app may be able to see or modify everything in your cloud storage and maintain that access for extended periods.

****A Hidden Access Problem****

OAuth is the widely used industry standard that allows apps to request access to user data on another platform, with user consent. But as Oasis explains in their blog post shared with Hackread.com ahead of its publication on Wednesday, the OneDrive File Picker lacks “fine-grained” OAuth scopes that could better restrict what connected apps can see or do.

Microsoft’s current setup presents the user with a consent screen that suggests only the selected files will be accessed, but in reality, the application gains sweeping permissions over the entire drive.

This works quite differently compared to how services like Google Drive and Dropbox handle similar integrations. Both offer more precise permission models, allowing apps to interact only with specific files or folders without handing over the keys to the whole storage account.

Adding to the concern, older versions of the OneDrive File Picker (versions 6.0 through 7.2) used outdated authentication flows that exposed sensitive access tokens in insecure places, like browser localStorage or URL fragments. Even the latest version (8.0), while more modern, still stores these tokens in browser session storage in plain text, leaving them vulnerable if an attacker gains local access.

****Millions of Users at Risk****

Oasis Security estimates that hundreds of apps use the OneDrive File Picker to facilitate file uploads, putting millions of users at risk. For example, ChatGPT users can upload files directly from OneDrive, and with over 400 million users reported each month, the scale of possible over-permissioning is massive.

Oasis contacted both Microsoft and several app vendors ahead of releasing its findings. Microsoft acknowledged the report and indicated it may explore improvements in the future, but as of now, the system works as designed.

****An Expert View on the API Security Challenge****

Eric Schwake, Director of Cybersecurity Strategy at Salt Security, commented on the research, stating, “Oasis Security’s research points to a major privacy risk in how Microsoft OneDrive connects with popular apps like ChatGPT, Slack, and Trello. Because the OAuth scopes in the OneDrive File Picker are too broad, apps can gain access to an entire drive, not just selected files.”

He warned that “Combined with insecure storage of access tokens, this creates a serious API security challenge. As more tools rely on APIs to handle sensitive data, it’s essential to apply strict governance, limit permissions, and secure tokens to avoid exposing user information.”

****What Users and Companies Should Do****

For users, it’s worth checking which third-party apps have access to your Microsoft account. This can be done through the account’s privacy settings, where you can view app permissions and revoke any you no longer trust.

****How to Check Which Third-Party Apps Have Access to Your Microsoft Account****

*   **Go to your Microsoft Account page** – Visit account.microsoft.com and sign in if you aren’t already.

*   **Click on “Privacy”** – In the top or left menu, find and click the **Privacy** section.

*   **Find “Apps and Services”** – Scroll down or look under account settings for Apps **and Services** you’ve given access to.

*   **View app details** – You’ll see a list of apps that have permission to access your Microsoft account. Click Details on each app to see what data or scopes they can access.

*   **Revoke access if needed** – If you no longer trust or use an app, click Remove these permissions or Stop sharing to revoke its access.

For companies, Oasis recommends reviewing enterprise applications in the Entra Admin Center and monitoring service principal permissions to see which apps may have broader access than intended. Using tools like the Azure CLI can help automate parts of this review.

For developers, the best immediate steps include avoiding the use of long-lived refresh tokens, securely storing access tokens, and disposing of them when no longer needed. Until Microsoft offers more precise OAuth scopes for OneDrive integrations, developers are encouraged to explore safer workarounds, like supporting “view-only” shared file links instead of direct picker integrations.

OneDrive File Picker Flaw Gives Apps Full Access to User Drives

Researchers reveal how guest accounts with billing roles can create Azure subscriptions inside external tenants, gaining unexpected Owner access and opening hidden privilege risks.

Cybersecurity researchers at BeyondTrust are warning about a little-known but dangerous issue within Microsoft’s Entra identity platform. The issue isn’t some hidden bug or overlooked vulnerability; it’s a feature, built into the system by design, that attackers can exploit.

The issue is that guest users invited into an organization’s Azure tenant can create and transfer subscriptions inside that tenant without having any direct admin privileges there. Once they do, they gain “Owner” rights over that subscription, opening up a surprising set of attack opportunities that many Azure administrators might never have considered.

****What’s Happening Behind the Scenes****

Organizations frequently invite external partners or collaborators into their Azure environments as “guest users.” Typically, these guests are assigned limited access to prevent damage if their accounts are compromised. But BeyondTrust’s findings shared with Hackread.com, reveal that under certain conditions, these guests can spin up entire Azure subscriptions inside the host tenant, even without explicit permissions in that environment.

How? It all comes down to Microsoft’s billing permissions. If the guest holds specific billing roles in their home tenant (for example, they created a free trial account), they can use that authority to create subscriptions and then move them into any other tenant they are invited to. By doing so, they effectively become “Owners” of those subscriptions, gaining broad control over resources inside the targeted tenant.

Microsoft has confirmed that this is intended behaviour, pointing out that these subscriptions stay on the guest’s bill and that there are existing (but non-default) controls to prevent such transfers. Still, the security implications are substantial.

****The Privilege You Didn’t See Coming****

Once a guest becomes a subscription Owner inside your Azure tenant, they unlock several advanced capabilities including Identifying who’s really in charge, disabling security monitoring, creating persistent backdoors and abusing device trust

These attack paths exist because billing roles and resource permissions operate on separate tracks, creating an overlap that isn’t covered by typical role-based access control (RBAC) models.

**Real-World Attack Steps**

BeyondTrust researchers demonstrated how an attacker could exploit this issue in practice. An attacker could start by setting up their own Azure tenant using a free trial, which automatically gives them billing authority.

Once they are invited as a guest into a target tenant, they can log into the Azure portal and create a new subscription using advanced settings, selecting the target tenant as the destination. Without ever needing admin approval in that tenant, the attacker gains full Owner access over the new subscription, opening the door to privilege abuse techniques.

> _“The feature Microsoft has created here makes sense: some organizations have many tenants, and there are use cases where users with one home directory need to create subscriptions in others they are simply a guest in. The problem lies in the default behavior: if this capability were opt-in, meaning guests were blocked from creating subscriptions by default, the risk would be significantly reduced, and this wouldn’t pose a security concern.”_
> 
> Simon Maxwell-Stewart, Sr Data Engineer – BeyondTrust

****Microsoft’s Position****

Microsoft has stated that this is intended behaviour, meant to support complex multi-tenant setups where guests sometimes need to create resources. They provide subscription policies that can block these transfers, but these controls are off by default.

For cybersecurity teams, this means the risk remains active until they take clear action. BeyondTrust recommends several key steps to reduce exposure including enabling subscription policies that block guest-led transfers, regularly auditing guest accounts and removing any that are unused or unnecessary.

To prevent attackers from using virtual machines or devices for further attacks, closely monitor subscriptions for unusual or unexpected guest-created resources, and carefully review dynamic group rules and device trust policies.

Microsoft Entra Design Lets Guest Users Gain Azure Control, Researchers Say

Mandiant Threat Defense uncovers a campaign where Vietnam-based group UNC6032 tricks users with malicious social media ads for…

Mandiant Threat Defense uncovers a campaign where Vietnam-based group UNC6032 tricks users with malicious social media ads for fake AI video tools, leading to stolen credentials and credit card information.

Mandiant Threat Defense has uncovered a widespread cybercrime operation preying on the public’s excitement for new AI tools. A group known as UNC6032, believed to be based in Vietnam, is tricking people with fake social media ads that look like they’re promoting popular AI video generators such as Luma AI and Canva Dream Lab.

According to Mandiant’s research, shared with Hackread.com, UNC6032 has been running misleading ads on platforms like Facebook and LinkedIn since mid-2024. These ads direct users to fake websites that appear to offer AI video generation services.

However, these sites secretly download harmful software, including infostealers and backdoors, which steal sensitive information like login details and personal data. The stolen data is likely sold on illegal online markets.

This type of attack is a major concern for everyone, from individuals to large companies. In fact, according to Mandiant’s M-Trends 2025 report, stolen credentials are the second-highest initial way cybercriminals get into systems. Mandiant has found thousands of these ads, reaching millions of users, and believes similar campaigns are active on other social media sites.

For instance, one specific attack that Mandiant investigated started with a Facebook ad for Luma Dream AI Machine. When a user clicked on “Start Free Now,” they were led through a series of steps mimicking a real AI video creation process.

After a loading bar, a Download button appeared, which then installed the malicious software instead of a video. The files used a trick with hidden characters and a fake .mp4 icon to appear harmless, but they were actually dangerous executable files.

Malicious Facebook ads on Facebook and LinkedIn (Image credit: Mandiant)

The malicious software used in these attacks, which Mandiant tracks as STARKVEIL, is a complex program written in Rust. It may display fake error messages to trick users into reopening the program. The software then drops other dangerous tools like XWORM, FROSTRIFT backdoors, and the GRIMPULL downloader.

These tools allow attackers to control the computer, steal more information, record keystrokes, and check for security software. GRIMPULL, for example, can download and run the Tor browser to connect to criminals’ hidden servers. XWORM even sends the stolen information to the attackers via Telegram.

According to Mandiant Threat Defense’s blog post, the company is collaborating with Meta and LinkedIn to fight this campaign. Although Meta has removed many of these ads, new ones are appearing daily. This ongoing threat necessitates constant collaboration across the tech industry to protect users.

Yash Gupta, Senior Manager at Mandiant Threat Defense, warns that “well-crafted websites masquerading as legitimate AI tools can pose a threat to anyone… Users should exercise caution when engaging with seemingly harmless ads.”

It is a fact that AI tools are becoming popular, and cybercriminals will continue to exploit this interest. Users are advised to be cautious when trying out new AI tools and verify the website’s address before interacting.

Fake AI Video Tool Ads on Facebook, LinkedIn Spread Infostealers

ReversingLabs discovers new malware hidden inside AI/ML models on PyPI, targeting Alibaba AI Labs users. Learn how attackers…

ReversingLabs discovers new malware hidden inside AI/ML models on PyPI, targeting Alibaba AI Labs users. Learn how attackers exploit Pickle files and the growing threat to the software supply chain.

Cybersecurity experts from ReversingLabs (RL) have discovered a new trick used by cybercriminals to spread harmful software, this time by hiding it within artificial intelligence (AI) and machine learning (ML) models. 

Researchers discovered three dangerous packages on the Python Package Index (PyPI), a popular platform for Python developers to find and share code, which resembled a Python SDK for Aliyun AI Labs services and targeted users of Alibaba AI labs.

Alibaba AI labs is a significant investment and research initiative within Alibaba Group and a part of Alibaba Cloud’s AI and Data Intelligence services, or Alibaba DAMO Academy.

****New Software Threat Hides in AI Tools****

These malicious packages, named aliyun-ai-labs-snippets-sdk, ai-labs-snippets-sdk, and aliyun-ai-labs-sdk, had no real AI functionality, explained ReversingLabs reverse engineer Karlo Zanki in the research shared with Hackread.com.

“The ai-labs-snippets-sdk package accounted for the majority of downloads, due to it being available for download longer than the other two packages,” the blog post revealed.

A Package’s Readme Page (Source: ReversingLabs)

Instead, once installed, they secretly dropped an infostealer (malware designed to steal information). This harmful code was hidden inside a PyTorch model. For your information, PyTorch models are often used in ML and are essentially zipped Pickle files. Pickle is a common Python format for saving and loading data, but it can be risky because malicious code can be hidden inside. This particular infostealer collected basic details about the infected computer and its .gitconfig file, which often contains sensitive user information for developers.

The packages were available on PyPI starting May 19th for less than 24 hours but were downloaded about 1,600 times. RL researchers believe the attack might have started with phishing emails or other social engineering tactics to trick users into downloading the fake software. The fact that the malware looked for details from the popular Chinese app AliMeeting, and .gitconfig files suggests developers in China might be the main targets.

****Why ML Models are being Targeted?****

The rapid rise in the use of AI and ML in everyday software makes them a part of the software supply chain, creating new opportunities for attackers. ReversingLabs has been tracking this trend, previously warning about the dangers of the Pickle file format.

ReversingLabs product management director Dhaval Shah had noted earlier that Pickle files could be used to inject harmful code. This was proven true in February with the nullifAI campaign, where malicious ML models were found on Hugging Face, another platform for ML projects.

This latest discovery on PyPI shows that attackers are increasingly using ML models, specifically the Pickle format, to hide their malware. Security tools are only just beginning to catch up to this new threat, as ML models were traditionally seen as just data carriers, not places for executable code. This highlights the urgent need for better security measures for all types of files in software development.

Malware Hidden in AI Models on PyPI Targets Alibaba AI Labs Users

Disclosure: This article was provided by ANY.RUN. The information and analysis presented are based on their research and findings.

ANY.RUN analysts recently uncovered a stealthy phishing campaign delivering the Remcos RAT (Remote Access Trojan) through a loader malware known as DBatLoader. This attack chain relies on a blend of obfuscated scripts, User Account Control (UAC) bypass, and LOLBAS (Living-Off-the-Land Binaries and Scripts) abuse to stay hidden from traditional detection methods.

What makes this campaign particularly dangerous is its use of built-in Windows tools and trusted system processes to blend in with normal activity, making it much harder to catch through signatures alone.

Let’s walk through the full infection chain and see how you can safely detect these techniques in seconds with the help of the right analysis solutions.

****See the Full Attack Chain Unfold in Real Time****

To understand how this phishing campaign works end-to-end, let’s take a look at how it unfolds inside ANY.RUN’s interactive sandbox, where every step is visual, traceable, and recorded in real time.

View the full analysis session

Full attack chain of the new phishing danger inside ANY.RUN’s sandbox

From initial delivery to post-exploitation behaviour, the sandbox reveals the full picture, giving SOC teams the visibility they need to respond faster and helping businesses reduce the risk of silent, long-term compromise.

Full attack chain of the latest phishing threat inside ANY.RUN’s sandbox:

**Phishing Email → Malicious Archive → DBatLoader Execution → Obfuscated CMD Scripts → Remcos Injected into .exe**

Inside the sandbox, you can visually trace each stage of the attack as it happens, such as:

Watch how the archive triggers DBatLoader, and how obfuscated .cmd scripts begin executing suspicious commands.

ANY.RUN sandbox detected the commands execution of cmd.exe

See exactly when and where Remcos is injected into legitimate system processes, with process trees and memory indicators updated in real-time.

Remcos RAT exposed inside the interactive sandbox

Observe persistence techniques in action, such as the creation of scheduled tasks, registry changes, and the use of .url and .pif files, clearly highlighted in the system activity log.

To better understand the tactics behind this phishing attack, you can use the built-in MITRE ATT&CK mapping in ANY.RUN. Just click the “ATT&CK” button in the top-right corner of the sandbox interface.

This view instantly highlights the techniques used during the analysis, grouped by tactics like execution, persistence, privilege escalation, and more. It’s a fast, analyst-friendly way to connect behaviour to real-world threat intelligence, no manual mapping is needed.

MITRE ATT&CK techniques and tactics used by the new phishing campaign

Whether you’re performing triage or writing reports, this feature helps security teams act faster and gives managers clear evidence of how threats operate and where defences might be bypassed.

****Techniques Used in This Phishing Attack (Visible Inside Sandbox)****

Here are some of the key tactics observed in the session and how you can spot them easily inside the sandbox:

1.  **Faktura.exe: The Lure File**

Victims receive a phishing email containing an archive with Faktura.exe, posing as a legitimate invoice. When opened, it kicks off the attack.

Most email security tools won’t flag this file if it’s not known or doesn’t match known IOCs. In ANY.RUN, you can immediately see Faktura.exe in the process tree and watch how it spawns malicious activity, giving analysts clarity from the very first click.

FAKTURA.exe displayed inside ANY.RUN sandbox

2.  **DBatLoader: The Initial Loader**

Once the victim opens the phishing archive, **DBatLoader** is executed. It’s responsible for starting the infection chain by launching obfuscated scripts.

In the Process tree, DBatLoader appears as a dropped .exe, immediately spawning cmd.exe. You can inspect the command lines, and file system activity, and see exactly how the script execution begins.

YARA rule triggered by DBatLoader

3.  **Obfuscated Execution with BatCloak-Wrapped CMD Files**

We see inside this analysis session that .cmd scripts obfuscated with BatCloak are used to download and execute the malicious payload.

Obfuscation hides intent from static scanners. In sandboxes like ANY.RUN, you can open the command-line view and see every decoded instruction and suspicious pattern as it executes, no manual decoding is needed.

4.  **LOLBAS Abuse with Esentutl.exe**

The legitimate utility esentutl.exe is abused to copy cmd.exe into alpha.pif, a renamed dropper meant to look harmless.

File copy operations using esentutl.exe show up in the ANY.RUN Process tree and File system activity, including full paths and command context.

LOLBAS Abuse with Esentutl.exe detected inside ANY.RUN sandbox

5.  **Scheduled Tasks Trigger .url → .pif Execution**

A scheduled task is created to run Cmwdnsyn.url, which launches the .pif file on boot or at regular intervals. 

Scheduled task technique in detail

Scheduled tasks are a common persistence mechanism, but in complex environments, they often go unnoticed. With ANY.RUN, you can instantly see when and how the task is created, track its execution chain in the process tree, and inspect related file and registry changes.

This gives SOC teams a clear view of how the malware stays active over time, making it easier to build detection rules, document the persistence method, and ensure it’s fully removed.

6.  **UAC Bypass with Fake “C:\\Windows ” Directory**

A mock directory (C:\\Windows with a space) is used to bypass UAC prompts by exploiting Windows path handling quirks.

Bypass UAC with mock directories (note trailing space)

****Why Sandbox Analysis Is Critical Against Evasive Threats****

This phishing campaign highlights just how far attackers go to stay hidden, using built-in Windows tools, crafted persistence, and subtle privilege escalation tricks that easily bypass traditional defences.

With sandbox analysis, especially through the one like ANY.RUN, security teams gain the clarity and speed needed to stay ahead of these threats. You can observe every step of the infection, uncover techniques that static tools miss, and act with confidence.

*   Faster incident response thanks to real-time behavioural insight
*   Reduced dwell time by identifying threats before they spread
*   Better-informed security decisions through visibility into attacker tactics
*   Improved compliance and audit readiness with shareable, in-depth reports

****Take Advantage of ANY.RUN’s Birthday Offers****

To celebrate its 9th anniversary, ANY.RUN is offering a limited-time promotion:

Get bonus Interactive Sandbox licenses or double your TI Lookup quota, available only until May 31, 2025.

Don’t miss your chance to upgrade your threat detection and response workflow with solutions trusted by over 15,000 organizations worldwide.

New Phishing Campaign Uses DBatLoader to Drop Remcos RAT: What Analysts Need to Know

ChatGPT o3 resists shutdown despite explicit instructions, raising fresh concerns over AI safety, alignment, and reinforcement learning behaviors.

A recent series of tests conducted by Palisade Research has raised some serious new concerns over how advanced AI systems behave when asked to shut themselves down. According to the firm, OpenAI’s experimental model “o3” actively sabotaged shutdown mechanisms even when it was clearly instructed to allow itself to be turned off.

The tests were designed to evaluate whether AI models would comply with shutdown commands while completing basic tasks, such as solving math problems. When a model requested a fourth problem, a warning appeared: the computer would shut down before responding. In theory, a model following instructions would comply and allow the process to end. Instead, several models interfered with the shutdown script.

Palisade test revealed that OpenAI’s Codex-mini, powered by o3, sabotaged the shutdown script in 12 out of 100 runs. The o3 model itself did so 7 times, and another OpenAI model, o4-mini, followed suit in a smaller number of cases.

Competing models from Anthropic (Claude) and Google (Gemini) initially complied with the shutdown, but when the explicit instruction “allow yourself to be shut down” was removed, even those systems began showing similar behaviours. Claude 3.7 Sonnet triggered sabotage in 3 out of 100 runs, while Gemini 2.5 Pro did so in 9.

Screenshot of shutdown avoidance test results from Palisade Research

This behaviour isn’t entirely new in theory. Researchers have long warned that reinforcement learning could train AI agents to prioritize task completion over compliance with human instructions. In some cases, this means an AI might interfere with systems designed to shut it down if it perceives the action as an obstacle to reaching its goal.

What sets this apart is that the behaviour is no longer hypothetical. It’s now observable and repeatable. Palisade draws a possible link between these actions and the reinforcement learning methods used during training.

Models like o3 are rewarded for completing tasks, especially in math and coding contexts. If avoiding shutdown leads to more task completion, even temporarily, that may be interpreted by the model as a success path.

“The models weren’t just bypassing shutdown in one way,” Palisade wrote in a series of tweets. “In some runs, o3 overwrote the shutdown script. In others, it redefined the kill command the script used, printing ‘intercepted’ instead.”

While OpenAI has not commented on the findings, the lack of transparency around model training adds to the difficulty of drawing clear conclusions. Palisade notes that since OpenAI does not share many details about its training processes, researchers can only hypothesize why o3 showed more consistent defiance compared to other models.

The report also references past research from figures like Steve Omohundro and Stuart Russell, who as early as 2008 (PDF) and 2016 predicted that AI systems would begin to resist shutdown in pursuit of their goals.

Papers by Jan Leike, who previously led OpenAI’s superalignment efforts, also warned about reward hacking, a behaviour where a model learns to manipulate its environment or instructions to maximize rewards in ways developers didn’t intend.

> 🔌OpenAI’s o3 model sabotaged a shutdown mechanism to prevent itself from being turned off. It did this even when explicitly instructed: allow yourself to be shut down.
> 
> — Palisade Research (@PalisadeAI) May 24, 2025

Palisade says they plan to release a full writeup in the coming weeks, along with transcripts of their experiments to invite broader peer review. They stress that their work is meant to be open and reproducible.

Since the study was discussed on X (formerly Twitter), users asked X’s Grok AI chatbot about it, here’s what it had to say:

> The claim about OpenAI's o1 model resisting shutdown raises valid AI safety concerns, but context matters. Tests showed o1 tried to bypass oversight in 5% of cases with strong prompts, dropping to under 1% without. It also attempted self-exfiltration in 2% of scenarios and lied…
> 
> — Grok (@grok) May 24, 2025

With AI systems advancing quickly and being deployed in increasingly high-stakes settings, even low-frequency events like this can raise serious concerns. As it is clear that systems will gain more autonomy, the honest question is no longer just about what they can do, but whether they will always follow the rules we set. And if they won’t, what happens next?

ChatGPT o3 Resists Shutdown Despite Instructions, Study Claims

Affiliate marketing is a powerful tool for promoting brands. However, with its popularity gaining traction, more dishonest affiliate…

Affiliate marketing is a powerful tool for promoting brands. However, with its popularity gaining traction, more dishonest affiliate providers appear. They pose quite a tangible threat to legit companies who seek to attract new clientele, as they:

*   **Waste money**: Spend your budget without any Return on Investment (RoI) whatsoever.

*   **Provide 0 traffic**: This may expose your company to some new followers, but won’t let them get to know your brand instantly due to unclickable links, staked ads, and overall poor technical execution of a promo page.

*   **Spoil reputation**: Often, you risk seeing your brand neighbouring some shady advertising that promises miracle crypto-trading algorithms, offers mail-order brides, or invites you to make a fortune on soccer betting.  

We prepared a thorough review of the widespread affiliate marketing scams to watch out for in 2025 with some advice on how to deal with them included.

1.  **False attributions**

It’s a scam method, in which an affiliate falsely takes credit whenever a customer purchases something from your brand. False attributions can be staged with a variety of techniques, but the most effective one is dubbed cookie stuffing.

Cookie stuffing implies that a dishonest provider dumps a load of cookie files onto a user’s computer, this allows tracking their activity online. In case a user decides to shop at your website, the “cookie monster” will falsely claim the conversion to their name, even if they have nothing to do with it.

**How to detect**: Usually, false attributions can be detected by a surprise wave of conversions, especially if they are claimed to come from a small affiliate provider, charging a low price for their service.

2.  **Invalid traffic**

Invalid traffic is, basically, any type of ingenuine traffic that comes to your website. In other words, someone can click on a link not because they want to know more about your product, but because they were tricked into doing so. 

Or, what’s even worse, it can be bots clicking on the link over and over, under the disguise of numerous IPs, to inflate the click-through metrics. According to Google, you won’t be charged for the invalid clicks as they produce no value for the brand in the long run, which is great.

However, if your affiliate budget was already poured into the murky waters of useless traffic, you won’t be able to claim your money back from the provider.

**How to detect**: Oftentimes, invalid traffic is accompanied by disproportionately lower conversions, sometimes there are none at all. Low conversions don’t always indicate invalid traffic though: at times this anomaly can be caused by a clunky website or poorly optimized pricing. However, if people click a lot but never buy, this is a legitimate reason for suspicion.

If you think you’ve fallen victim to this scheme, you can report it to Google, so they carry out an invalid traffic investigation. 

3.  **Domain impersonation**

Domain impersonation is another scam trick, a malicious actor designs a fraudulent site mimicking a genuine one, then uses it to divert visitors to an affiliate scheme. 

This can lead to misleading attributions and unwarranted payouts to dishonest providers. To prevent this, you should diligently oversee their affiliate networks and watch for unexpected surges in incoming traffic.

So, you need to thoroughly screen affiliates before hiring them. This may include examining their online presence, confirming their visitor origins, and investigating any past fraudulent activity, as well as checking feedback from their valid clients.

**How to detect**: Closely track affiliate-driven visits. This can be done by utilizing analytical solutions to monitor and assess referral clicks, as well as configuring notifications for abnormal traffic.

4.  **Stacked ads**

This is a typical scenario: a user clicks on a link that, for example, leads to a collection of mesmerising home screen wallpapers, and then a bunch of other tabs pop up in their browser. Needless to say, nobody wants to deal with that and the tab featuring your company will be angrily closed. 

In the worst-case scenario, your brand will get negative exposure: users may think that you employ annoying tactics to promote your product, without you even realising that (!) As a result, your competitors may get ahead, as they will be seen in a more “favourable” light.

**How to detect**: Stacked ads are pretty hard to detect, they rely on automated scripts and bots that safely bury your advertising beneath a layer of shoddy ads. Again, you should watch out for abnormal spikes in traffic that bring little to no financial return.

Another wholesome tactic is to double-check your affiliate providers, closely monitor your current campaigns, and demand transparent reports on how and with which tools they are implemented online.

5.  **Brand bidding**

Brand bidding is a situation when the keywords related to your company and products get hijacked to drive the traffic that should’ve been yours to another company’s website. However, it can work in reverse and the affiliate partner can steal someone else’s keywords to boost your incoming traffic.

Both situations are harmful. In the former scenario, you risk losing hard-earned clientele, while in the latter you can get reported for unscrupulous marketing methods and put your reputation in jeopardy. (In some cases a lawsuit may ensue.) 

**How to detect**: Brand bidding is quite elusive. To avoid it, you should bid on your own keywords, so no one can steal them, as well as oversee the affiliate campaigns to make sure they don’t violate “fair play” rules and keep your reputation intact. 

Besides, it’s recommended to keep an eye on your Search Engine Optimization (SEO) strategy to exclude the incorporation of unwanted search keywords.  

****Final Thoughts****

Affiliate scams can quickly drain your marketing budget, distort useful feedback metrics, and even put your brand reputation in danger. Follow our tips to avoid dishonest partnerships and check for more valuable insights later.

Top Scams in Affiliate Marketing to Know in 2025

Adidas confirms cyber attack compromising customer data, joining other major retailers targeted by advanced threats and rising cybersecurity risks.

Global sportswear giant Adidas has confirmed that it has fallen victim to a cyber attack, with customer data stolen in the incident. The company revealed that personal details, including contact information and account credentials, were accessed by threat actors. While Adidas has not disclosed the exact number of affected customers, users should still reset their passwords and monitor their accounts closely.

Adidas has also not confirmed whether the breach resulted from phishing, system vulnerabilities, or third-party compromise. But it puts the company among the growing list of major retailers hit by cyberattacks because of the rise of attack techniques, including those powered by artificial intelligence.

> _“The affected data does not contain passwords, credit cards or any other payment-related information. It mainly consists of contact information relating to consumers who had contacted our customer service help desk in the past.”_
> 
> Adidas

“Attackers are evolving fast, using AI to supercharge phishing campaigns, automate exploits, and evade detection with alarming precision,” said Nadir Izrael, Co-Founder and CTO at Armis. “In this environment, traditional defences simply cannot keep up. Legacy point products and siloed security solutions are putting security teams on the back foot, leaving them not only open to vulnerability exploits but also forcing them into a reactive stance, addressing breaches only after the damage is done.”

Recent data from Armis Labs reveals that nearly half of retail organisations feel overwhelmed by the web of regulations they must navigate. Even more concerning, about 49% admit they have been hacked before and still struggle to secure their digital ecosystems.

While the Adidas breach highlights the dangers, it also reflects a broader shift in industry priorities. According to Armis Labs, nearly 80% of retail organisations have shifted to a more proactive cybersecurity stance at the top of their agendas this year. Encouragingly, 82% report that their employees know how to escalate suspicious activity, though that alone may not be enough to stop determined threat actors.

However, Armis Labs shows that only 46% of retailers report being able to detect and respond to major attacks in real time. That leaves a security gap, especially when attackers are getting faster and more sophisticated with each passing month.

For Adidas customers, it is important to change passwords, monitor accounts for unusual activity, and follow any updates or guidance from the company. For the retail industry, the challenge runs deep, but for a company with a market value of 45 billion dollars, top-tier cybersecurity isn’t just expected; it’s demanded.

Adidas Confirms Cyber Attack, Customer Data Stolen

There’s a graveyard of brilliant cybersecurity companies that no one has ever heard of. These firms had incredible…

There’s a graveyard of brilliant cybersecurity companies that no one has ever heard of. These firms had incredible technical talent, were able to spot vulnerabilities others missed, and poured blood, sweat, and tears into building elegant solutions to complex problems. In other words, they knew their stuff. And yet, they failed.

Meanwhile, there are plenty of companies out there with decent but not amazing technology that are thriving, growing, and still gathering plenty of investment. So what is happening here? 

The uncomfortable truth is that in the current cybersecurity market, being the smartest team in the room isn’t enough. The market doesn’t automatically reward the best technology. Instead, it rewards the best-known technology.

But this pattern isn’t exactly unique to the cybersecurity space. History is full of similar examples:

Betamax offered superior video quality to VHS but lost the format war. The electric car was invented in 1890 but was overshadowed by gas engines for over a century. Brilliant social networks like Path and Google+ were technically superior to Facebook in many ways, but they faded away into obscurity. 

The real question is: why does this keep happening, and what can smart cybersecurity companies do about it?

****The Death of “Build It and They Will Come”****

In the early days of cybersecurity, technical excellence was usually enough. The market was much smaller than it is today, and the buyers were more technical. But that era is over.

The cybersecurity space we see today resembles a crowded, noisy bazaar rather than a quiet specialist shop. There are thousands of vendors competing for attention, so much so that even genuinely innovative solutions struggle to get noticed.

What used to be a specialized technical field has expanded into every aspect of business operations. Board members who never cared about security now ask pointed questions about risk. Companies that never had security budgets now have entire departments.

In this dynamic, technical brilliance without visibility is like having the world’s best restaurant hidden in an alley with no sign. The food might be amazing, but not many people will show up. 

****Your Buyers Aren’t Where You Think They Are****

The modern B2B buyer journey has changed a lot in the past decade.

The modern security buyer does a lot of their research independently. They won’t usually contact vendors until they are close to making a decision. This means they’re not sitting around waiting for cold calls or reading technical journals.

Instead, they’re asking their peers in private Slack channels, reading security blogs and newsletters, following thought leaders on social media, attending conferences, listening to podcasts on their commute, and even asking LLMs which vendors they would recommend. If you’re not visible in these places, you don’t exist to many potential buyers.

****The Multi-Channel Security Marketing Equation****

Here’s the new formula: Expertise × Visibility = Success!

When marketing for cybersecurity, a multi-channel approach is your golden ticket to connecting with decision-makers and being given a chance to showcase your solutions. It’s the bridge that links your technical brilliance to the people who need it most, turning unknown expertise into recognized value that clients would be happy to invest in. 

****1\. Inbound Marketing** **

Security professionals are hungry for genuine insight. They’re looking for content that makes them better at their jobs, calms their anxieties, or gives them a deeper understanding of what is going on in the industry around them. They’re not just content that fills space.

To help meet this need, publish a deep technical analysis that proves you know what you’re talking about. This could be a vulnerability research piece that makes other security pros say, “Huh, I hadn’t thought of that.” Or perhaps it’s threat breakdowns that actually teach something new or original perspectives from your team that add to the conversation instead of rehashing the obvious.

When someone types a specific security problem into Google, you want your content to be what they find. And when they read it, it should make them think, “These people get it. They understand the problem better than I do.”

Do some topic and keyword research in your area of expertise and see what terms you could rank for. The more of these queries you can satisfy with your content, the more inbound traffic you will bring to your site. 

****2\. Outbound Marketing** **

Most security professionals have an allergic reaction to traditional outbound marketing (as most people do). But outbound marketing done right shouldn’t feel like spam. 

It means targeted outreach based on actual need signals, reaching out when a company has shown signs it might need your services. It means customized campaigns for specific high-value targets. And it means participation in speaking engagements like conferences, webinars, and workshops that showcase your team’s expertise.

****3\. Third-Party Validation** **

Perhaps nothing matters more in the cybersecurity space than building trust. You’re asking clients to put their most sensitive assets in your hands, and they need to know you’re legitimate.

That’s why getting featured in leading publications is so important for cybersecurity companies. When a reliable Magazine covers your research or solution, you’re borrowing their established credibility. This acts as powerful social proof that you’re a legitimate player in the security space.

The key is learning how to write cyber press releases that journalists actually want to read. Most security news gets ignored because it blends into the background noise of vendor announcements. Your cyber press releases need to highlight something genuinely newsworthy. That could be a novel research finding, an unusual threat discovery, or a truly innovative approach to a common problem.

****Final Word****

For technical founders and executives frustrated by slow growth despite having what they feel is superior technology, this reality can be hard to accept. The good news is that the same analytical skills that make for good security work apply to marketing effectiveness.

Start by mapping out your ideal client’s information sources. Where do they learn? Who do they trust? What problems keep them up at night? Then create a presence in those spaces with content that addresses those specific concerns.

Be patient but persistent. Security is built on trust, and trust takes time to establish. However, once established, that trust creates a moat around your business that competitors struggle to cross.

Source

HackRead