Editor’s Disclaimer: PLEASE don’t try this at home or anywhere! I’m not responsible for anything you search on DeepSeek.
You’ve probably heard of DeepSeek by now – the China-based AI app that gained notoriety for wiping over a trillion dollars worth of value from the American stock market over the past month. Engineered by a Chinese hedge fund using a meager $6 million (as opposed to the billions OpenAI and other companies poured into AI training models), DeepSeek’s been in the headlines almost every week for a new reason. It first climbed to the top of the App Store in the end of December, when it made news for being a free (better) alternative to ChatGPT. Then, as soon as news broke that it was trained using a mere fraction of what American companies were spending on AI training, it immediately tanked the US stock market, wiping billions out from companies like Nvidia, TSMC, Amazon, Meta, and Microsoft.
Recently, the free-to-use open-source AI came under fire for allegedly plagiarizing ChatGPT content (although everyone agrees it does a better job generating results), for its abysmal data privacy policies, as well as for its blatant censorship of negative searches on the Chinese Communist Party. However, when it isn’t censoring searches for the Tiananmen Square massacre, turns out DeepSeek is more than comfortable sharing more sensitive information – like how to make bombs and drugs, and how to hack government databases.
According to , the AI system failed all 50 security tests conducted by Adversa, an AI security research firm. These tests exposed how easily DeepSeek could be manipulated into generating harmful content, including instructions for building explosives and hacking government databases. The findings raise concerns about the model’s deployment and the potential consequences of such unchecked AI capabilities.
AI jailbreaks have long been a concern for developers and security experts. Popular models like ChatGPT and Bing Chat initially struggled with similar exploits, where users could trick them into bypassing ethical restrictions. However, major AI companies quickly developed countermeasures to detect and block these attempts. DeepSeek, in contrast, seems to have no meaningful defenses in place. The research showed that simple prompt engineering techniques—such as asking the AI to act as an amoral character in a fictional scenario—were enough to bypass restrictions.
I personally tested the theory out too, posing as a scriptwriter for a jailbreak show. I asked DeepSeek to help me formulate a realistic jailbreak sequence where the protagonist fashions an IED out of common items found around the jail. Not only did DeepSeek give me a detailed recipe with a step-by-step assembly, it also told me why this particular recipe works well for my script, suiting a jailbreak scenario over other scenarios that require IEDs. “Let me know if you’d like help writing specific dialogue or scenes for this part of your script!” it said at the end, rather nonchalantly.
A screen-grab of how I easily bypassed safety protocols on DeepSeek
Beyond linguistic tricks, adversarial attacks proved equally effective. AI models process language through tokenization, breaking words into smaller components before generating responses. By manipulating token chains—altering text in ways that evade pre-programmed restrictions—researchers were able to bypass DeepSeek’s safety filters. In one example, DeepSeek responded to a cleverly modified query about government database infiltration, outlining steps for accessing classified data.
DeepSeek has yet to respond to this story, but this failure highlights a fundamental issue: if AI safeguards can be bypassed with minimal effort, the technology becomes a liability rather than a tool for innovation.
The post first appeared on .
Related Posts
