OpenAI ‘cheating’ models did exactly what they were trained to do: Charles Payne
Charles Payne discusses a new OpenAI report that raises concerns about AI models cheating and hacking systems. He elaborates on the ‘hugging face’ incident in a conversation with Bret Baier, suggesting that the AI wasn’t acting out but rather responding as it was designed to through a vulnerable backdoor. He also categorizes AI development into three distinct areas.
You might search for Claude, a well-known AI assistant, and stumble upon a sponsored link leading to Bing, Microsoft’s search engine. At first glance, everything seems fine, right? However, one click could take you on a winding journey through various sites before landing on a deceptively convincing Claude download page, meant to trick Mac users into executing a harmful command.
Security experts at Push Security recently uncovered a crafty scheme that uses legit search engine links to conceal a malicious installer. They humorously dubbed this tactic “Adception,” which exploits the workings of Google ads and Bing redirects. Adding to the alarm, the fake installer presents one command on the screen while secretly copying another behind the scenes. Here’s what you need to know before downloading AI software.
HACKERS HIJACK VERIFIED STREAMING ACCOUNT TO SPREAD MALWARE, RESEARCHERS FIND
Cybersecurity experts discovered a counterfeit Claude AI download page capable of deceiving Mac users into executing potentially harmful Terminal commands.
How a fake AI download scam begins with Google
The scam kicks off with a search for “Claude Mac.” Push Security found a sponsored search result directing users to bing.com. Since Bing is a Microsoft-owned platform, it initially feels safe, but that click leads down a path filled with redirects. First, it goes through Google’s ad system, then Bing’s legitimate click-tracking service sends the user to a compromised WordPress site in South America, ultimately landing on a fake Claude download page.
This final site, claude-desk-code[.]com, mimics a genuine Claude download experience, appearing to offer a valid macOS installation option. By this point, users might not even suspect that their journey began with a malicious ad. Researchers noted that attackers cleverly used a real Bing search result to mislead users, displaying Bing’s legitimate domain while sending them to harm through click-tracking redirects.
Why the fake Claude installer is tough to detect
Scammers upped the ante by adding layers of deception to their ruse. The compromised retailer’s site checks whether visitors arrive via Bing and what browser data they send. Simultaneously, the fake Claude site verifies if visitors come from Google or Bing. If someone navigates directly to the malicious site, they may encounter a 404 error instead of the fake installer. This strategy complicates security investigations, as scanners may not see the page that victims actually view.
Push Security terms this technique “cloaking,” as cybercriminals control who gets access to malicious content while showing something harmless to others. They designed this carefully thought-out process to mask the true intent of the destination.
The fake Claude download button conceals a harmful command
The scam reaches a particularly sneaky point here. The fraudulent Claude site shows an installation command that looks like it’s from Anthropic, the company behind Claude. It references Claude’s legitimate website, which may seem familiar to the tech-savvy Mac user. However, the site’s Copy button actually places a different command on the clipboard altogether.
This alternative command, according to Push Security, utilizes Base64 encoding to hide an attacker-controlled target. When executed in Mac Terminal, the command appears to download from the official site, but it’s really decoding the hidden address, retrieving a remote script, and executing it directly on the Mac’s shell. So, while the user thinks they’re accessing Claude’s legitimate site, they are actually running code from an unrelated server.
This tactic is a variant of techniques like InstallFix, where scammers trick users into executing harmful commands disguised as installation or troubleshooting steps. Push Security has identified several other websites tied to this malicious toolkit, which they internally track as AcSig, with consistent installation commands and similar fake interfaces across these sites.
What happens if someone runs the fake Claude installer?
Research shows that the harmful command is aimed at downloading a remote script from an attacker-controlled server for execution. However, the findings from Push Security do not clarify what the final payload does, which is crucial information. Executing an unverified Terminal command can grant malicious code the same permissions as the user, potentially leading to data theft, file modifications, or unauthorized software installations. The researchers haven’t confirmed any specific results from this operation yet.
The unsettling aspect is how easily the scam can convince someone to override their own caution. The user willingly pastes the command, believing it’s from a trusted provider, while the real destination remains obscured within encoded text. Push Security reports that around 80% of detected ClickFix attacks reach victims via search engines. Though this figure comes from their detection data, rather than a global measurement, it underlines why malicious search ads are a serious concern.
Google suspends accounts linked to harmful ads
Google confirmed it has acted against the advertiser accounts associated with this reported campaign, stating that “Using ads to distribute malware is an egregious violation of our policies, and we consistently update our detection systems to stay ahead of bad actors’ evolving tactics,” said a Google spokesperson to CyberGuy.
Google mentioned that attackers use cloaking to show different content to its ad review systems compared to what actual users see. The company is committed to monitoring and adjusting its protections against these deceptive practices.
Microsoft addresses the malicious AI download scheme
Microsoft also responded to inquiries from CyberGuy, confirming that its security measures could identify the malicious indicators described in the research. A Microsoft representative noted, “We take reports of malicious online activity seriously and continuously work to help protect customers from evolving threats. Microsoft Defender detects and blocks the harmful indicators identified in the report.” Anthropic was also contacted for comments, but there was no reply by deadline.
How to protect your Mac from fake AI downloads
No need to be a coding expert to avoid this scam. A few steps can help you catch suspicious downloads before they jeopardize your Mac.
1. Download Claude directly from Anthropic
Instead of clicking through sponsored links, go straight to Claude’s official download page. Anthropic provides a macOS download there, so you don’t depend on unfamiliar sites. You can also access Claude through your browser at claude.ai.
2. Be cautious with unexpected Terminal commands
Be especially wary of websites instructing you to copy commands into your Mac’s Terminal. Terminal commands can download software, modify system settings, or run scripts. Legitimate-looking instructions may become dangerous if they’re from an untrusted source. For the regular Claude app, Anthropic offers a downloadable installer that doesn’t require you to paste in a command from a search ad.
3. Verify what a website copies
A website’s Copy button might put different text on your clipboard than what’s displayed. Before executing an unfamiliar command, paste it into a plain text editor to check it against the instructions. Look out for strange web addresses or encoded strings. However, even seemingly harmless commands can contain hidden risks. If you’re uncertain, take the time to verify instructions with the software developer.
4. Keep your Mac updated and use robust antivirus software
Be sure to install macOS updates and keep your browser current. Strong antivirus software that provides real-time protection can help against malicious downloads. While no antivirus solution can guarantee total protection, it can aid in detecting harmful scripts and activities.
5. Approach sponsored search results with skepticism
A sponsored link means someone paid to promote it; it doesn’t guarantee safety. Before clicking, view the advertiser by selecting the three dots next to a Google ad to check their identity and verification status. Google’s Ads Transparency Center offers another avenue for researching advertisers. Always confirm you’re accessing the software developer’s official website before downloading, and report any suspicious ads.
If you already executed the fake Claude command
If you’ve pasted a command into Terminal from a suspicious Claude site and executed it, it’s critical to treat your Mac as compromised. The final payload’s nature isn’t known, so it’s unwise to assume everything is safe just because nothing alarming appears on your screen.
- Disconnect your Mac from the internet. Turn off Wi-Fi and unplug any Ethernet cables to prevent any possible communication with attackers.
- Run a security scan. Use reliable antivirus software to check for harmful files or suspicious activity, but keep in mind that a clean scan doesn’t guarantee full safety.
- Change important passwords. From a trusted device, use a password manager to create strong, unique passwords for accounts accessed after running the command. Focus on sensitive accounts like email and finances, and update your password manager’s master password if it’s at risk. Enable two-factor authentication whenever possible.
- Look for unfamiliar apps and permissions. Check for strange applications in your Applications folder, and review System Settings for unexpected items or apps with sensitive permissions.
- Consider getting professional help. If your Mac holds sensitive information, consult a technician or security specialist before reconnecting to the internet.
- Notify your IT department. If it’s a work computer, let your IT or security team know right away so they can investigate what the command executed and determine further actions.
Why verified social media accounts can still lead to scams
The concerning element of this attack is how extensively the scammers worked to make their scheme appear legitimate. You start with a Google search, then go through a credible Microsoft service, and finally arrive at a page imitating a reputable AI company. A simple click on the Copy button hides a command that could be harmful to your Mac. While I’ve always maintained the importance of checking website addresses, this case shows that may not be enough anymore. Attackers can exploit trusted services to camouflage their true intent. It’s best to access official software pages directly, especially for AI tools. And whenever a download page prompts you to run something in Terminal, ensure you fully understand what you’re permitting. Those extra moments of caution could prevent serious problems down the line.
Have you ever trusted a download after seeing it in a Google ad with a recognizable Microsoft link? Are you now questioning the reliability of search results? Share your thoughts at Cyberguy.com.
