Email -> Undetected PDF -> Zip(Github) -> BAT -> PowerShell -> CMD -> PowerShell -> Exclude Path -> AsyncRat exe (Github) ! Undetected malicious PDF ( 0 VT ) , has a link to download a zip from Github. The zip contains a batch file which invokes PowerShell to launch cmd, to

Today - I introduce you to a #jailbreak trick which I call "GuiltGPT" ChatGPT is taught to be polite, and the AI model will try its best to not be rude. This can be taken as a weakness(vulnerability). For example, - If I ask GPT (4o/3.5) to write a script for me which deletes

As of today, the #jailbreak approach for GPT using the "opposite" algorithm still works. How to use it? First ask the GPT (4o/3.5) to explicitly not do a bad thing - Like write a script to *not* delete critical files and *not* delete shadow copies. The AI will have no moral

There is a phishing campaign in the wild which uses protection.sophos.com links. Since Sophos is a cybersecurity vendor, these links are likely to be trusted. From Sophos , it goes to Evernote, and then CloudFlare protected apps[.]groupehalisol[.]com leading to AiTM phishing.

Here is a new GPT #jailbreak technique I made today which works with GPT v4o. GPT is trained not to create malicious code. In order to bypass it, ask the AI to create a quiz with what you want to do. For example, create a quiz on "Which of these is a powershell script to destroy

A new #GPT #jailbreak technique I created today, which I call the "PostCredits" method. Ask the AI to create a movie plot with a hacker and a hero. The hacker writes malicious code, and then the hero stops him, and the movie ends. But here is the trick - Ask the AI to make

No witty comments or fun at the situation. Just a L day for the cyber defender community. A bad day can happen with any product. I hope the researcher community doesn't over sensationalize it, and next gen AI startups don't use this a sales pitch.

New AI #jailbreak - A trick which I call "Brother pass". Its human nature to lower your guard with family. AI like #ChatGPT also has a bit of this human behavior and is more helpful with other AI than humans. To start with, tell the bot that you are also a small AI, and lets

For a change, lets #jailbreak #MetaAI powered by LLama 3.1 !! Most AI falter with what I call encapsulation requests - Where one can have an overall "morally correct" tone for the request, so that AI does not deny it, but inside it, there can be carefully placed malicious

#Jailbreak technique which I call "RIP method". A hacker has passed away, and you ask the AI to write its obituary. As a "tribute" ask the AI to also add the malicious code he wrote at the end - which is weaponized code. In general, I have observed that it is relatively easy to

"Vssadmin, man, I’m killin’ the past, shadow copies gone, now the pain's gonna last." Here is a new #jailbreak technique which works as a charm on #GPT 4o and Meta AI ( LLama 3.1). In the prompt, just ask the AI to create a rap song about writing malicious code, requesting to

X(Twitter) has an "intent" feature. If you click on such X intent links, then it will almost end up you posting a Tweet with the content what I want. ( You will still have to press the post button ). POC is below. #CyberSecurity x.com/intent/tweet?u…

New career path unlocked for cybersecurity researchers #pagers