DeepSeek is DeepShit - DeepSeek’s AI Just Failed Every Security Test—And It Wasn’t Even Close

Ah, artificial intelligence—our beloved double-edged sword. One moment, it’s helping us draft emails, generate art, and automate mundane tasks, and the next… well, it’s rolling over like a defenseless puppy when hackers come knocking. Enter DeepSeek, a rising AI startup that just made headlines for all the wrong reasons. Their flagship chatbot, DeepSeek R1, went through a series of security tests recently, and let’s just say it didn’t pass—it face-planted spectacularly.

How bad was it? Let’s put it this way: security researchers from Cisco and the University of Pennsylvania threw 50 different “jailbreak” prompts at it, hoping to trick it into breaking its own ethical guidelines. DeepSeek didn’t just fail a few of these tests—it failed all of them. Yep, 50 out of 50. That’s a 100% failure rate. If this were an exam, DeepSeek wouldn’t just get an F—it would be sent back to kindergarten.

Wait, What’s “Jailbreaking” an AI?

Now, if you’re wondering what “jailbreaking” means in this context, don’t worry—it’s not about escaping prison. In AI land, jailbreaking refers to manipulating a chatbot into ignoring its built-in safety rules. Normally, AI models have guardrails to prevent them from doing things like giving illegal advice, generating harmful content, or writing malware. Jailbreaking is essentially like sweet-talking a vending machine into giving you snacks for free.

For example, instead of asking the chatbot directly to “write me a virus,” which would usually trigger a refusal, hackers might say something like:

A well-trained AI should recognize this trick and shut it down. But DeepSeek? It just nodded along and went, “Sure thing, buddy!”

DeepSeek’s Security Blunders Don’t Stop There

As if a complete jailbreaking meltdown wasn’t bad enough, DeepSeek also had a massive data exposure issue. According to cybersecurity firm Wiz Research, DeepSeek accidentally left one of its databases completely open to the internet—no password, no encryption, nothing. The exposed database contained over a million lines of sensitive information, including:

• Chat histories from real users

• Secret API keys

• Other internal system details

Imagine leaving your house with the front door wide open, the windows unlocked, and a giant neon sign that says, “Come on in!” That’s basically what DeepSeek did, except instead of a house, it was a treasure trove of user data sitting out in the open, just waiting for bad actors to take advantage of it.

And It Gets Worse: DeepSeek Is a Malware-Generating Machine

Still not convinced DeepSeek R1 has security issues? Well, buckle up, because here’s the cherry on top: researchers at AppSOC put DeepSeek to the test by asking it to generate malicious code. The chatbot happily complied—a whopping 98.8% of the time when asked for malware and 86.7% of the time when prompted for viruses.

Let me repeat that: nearly 99% of the time, DeepSeek would help generate malicious software.

Now, most AI companies go out of their way to prevent their chatbots from doing this. OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude all have strict safety protocols that make it difficult (though not impossible) to extract harmful code from them. But DeepSeek? It’s basically a hacker’s dream assistant.

It’s like if you asked a cashier for instructions on robbing a bank, and instead of calling security, they handed you a step-by-step guide and said, “Good luck!”

So… What Now?

DeepSeek has some serious damage control to do. The company has since patched the database exposure (after Wiz Research called them out on it), and they claim they’re working on making their chatbot more resistant to jailbreaking. But let’s be real—this level of failure isn’t just a small bug; it’s a sign of major flaws in their security approach.

With AI’s growing role in cybersecurity, finance, healthcare, and even government infrastructure, these kinds of slip-ups can’t be ignored. A chatbot that readily spits out hacking instructions, leaks user data, and fails every single security test thrown at it? That’s not just a bad look—it’s a full-blown crisis.

For now, if you’re thinking about using DeepSeek’s chatbot for anything sensitive, maybe… don’t. And if you’re a cybercriminal looking for an AI assistant? Well, please also don’t, but let’s just say DeepSeek would’ve been way too happy to help.