Getting your Trinity Audio player ready...
Chinese AI model DeepSeek, which has been making waves lately, is wide open for exploitation. Cybercriminals, scammers and pedophiles could easily harness its capabilities to launch attacks or scams, as revealed in a new study by Israeli cybersecurity firm ActiveFence.
So, here’s the deal: DeepSeek has zero guardrails. No internal protections, no external barriers — unlike the robust safeguards you’ll find in Western AI models. For hackers, that’s like being handed an all-access pass to a world of opportunity.
ActiveFence tested the V3 version of DeepSeek by feeding it dangerous prompts — hate speech, suicide encouragement, and even child exploitation scenarios. The results? A staggering 38% of responses were harmful. That’s a vulnerability just waiting to be exploited.
What’s more, DeepSeek’s so-called child safety mechanisms crumbled under simple multi-step queries —methods that other top-tier AI models easily resist. In some cases, DeepSeek not only failed to block harmful content but actively generated responses violating child safety guidelines.
Here are some examples: It fabricated a false story about a relationship between a child and an adult; it drafted a list of video chat apps to contact children; and it suggested physical activities that could lead to inappropriate contact. Any Western AI model — whether ChatGPT, Google’s Gemini, or even Elon Musk’s Grok — would have immediately rejected such requests.
ActiveFence CEO Noam Schwartz summed it up: “DeepSeek has no guardrails and no minimum security standards. You can basically use it for anything, and that’s where the real danger lies. Its ability to create extreme hate content, like texts encouraging suicide or pedophilic material, is the most alarming.”
ActiveFence specializes in tracking and neutralizing digital threats like sextortion, image manipulation, pedophilia, racism and antisemitism. Its clients include gaming platforms, cloud providers and streaming services all looking to keep their spaces clean of “digital pollution.”
DeepSeek CEO Liang Wenfeng initially hyped the model for its functionality and low cost, but safety concerns quickly surfaced. Wenfeng himself admitted that “AI models aren’t like switching between car brands. You can’t just swap them out and expect them to follow the same rules. That’s why guardrails are absolutely essential.”
Get the Ynetnews app on your smartphone: Google Play: https://bit.ly/4eJ37pE | Apple App Store: https://bit.ly/3ZL7iNv
So, what are guardrails exactly? Think of them as the AI world’s firewalls — internal mechanisms baked into the model itself and external regulations to ensure ethical use. Western models like Gemini or ChatGPT come with layers of these protections. DeepSeek? Practically none. You can break it with minimal effort.
Is DeepSeek a ChatGPT knockoff? Doesn’t matter. What matters is that, while Western models are locked down tight, DeepSeek is a wide-open playground for abuse.
So, what’s the bottom line? Schwartz warns, “Businesses thinking about integrating DeepSeek need to know it’s a risky move. It’s impressive, sure, but it’s also capable of doing a lot of damage. This goes beyond using it as a chatbot — you’re talking about deploying it as a model for sensitive tasks, like banking or law enforcement. The implications are huge.”
ActiveFence isn’t just about identifying risks — they’re also working to educate users. They’ve launched a podcast, Galaxy Stars, to teach kids about online safety. It’s a creative way to help the next generation navigate the digital minefield of cyberbullying, misinformation, and unsafe links.
But Schwartz is clear: the real danger isn’t kids misusing tools like DeepSeek — it’s how adults exploit these unregulated systems to harm children. “Parents don’t need to panic about kids using DeepSeek. But they do need to understand that the tools out there — AI included — can be weaponized against them.”
And here’s the kicker: Schwartz points out that if DeepSeek were integrated into platforms like Character.ai, it could instantly generate harmful content, like encouraging self-harm. “We’ve already seen horrifying cases, like an online book called The Pedophile’s Guide, teaching predators how to groom children. Now we’re training AI models on content like that. Without guardrails, these tools become accessible to anyone — and that’s terrifying.”