AI vs. AI: How Abnormal Fights DeepSeek AI-Powered Phishing Attacks

In news announced this week, Chinese AI company DeepSeek has developed some of the most advanced large language models seen in recent years, sparking both excitement and concern in the technology and cybersecurity sectors. Their latest models, DeepSeek-V3 and DeepSeek-R1, match models like OpenAI's GPT4o and O1 across multiple benchmarks.

Much like ChatGPT and other generative AI platforms, the release of DeepSeek will enable cybercriminals to create sophisticated attacks at scale. And if history repeats itself, we’re likely to see an influx of them in coming weeks as threat actors use this newly available technology to their advantage. But what makes DeepSeek different from other generative AI platforms?

First, DeepSeek was supposedly trained by a team based in China with relatively few computational resources. Models of this quality have thus far only come from extremely well-funded research labs (such as OpenAI). We should therefore expect to see more powerful models coming from less well-resourced groups in the future.

Second, DeepSeek is an open-source model (like Meta's Llama series), which means that anyone, anywhere in the world, can download the model, modify it at will, and run it on their own infrastructure. Unlike tools like ChatGPT, which have a company (OpenAI) in theory responsible for ensuring that people do not use the tool for malicious purposes, open source models can be utilized by cybercriminals with no guardrails.

As DeepSeek AI models gain global traction, they could bring significant security concerns, particularly when it comes to social engineering. These language models have the ability to easily compose a convincing phishing or business email compromise (BEC) attack with correct grammar and specific calls to action, making it even harder for individuals to distinguish between a safe communication and a malicious one.

For example, I asked DeepSeek to create a business email compromise attack, pretending to be a vendor who needed to update banking details, and it provided me with this copy:

While DeepSeek does explain that this is an example and should not be used for malicious purposes, it provides the content without any form of censorship. As you can see, the email is very well-written with no grammatical or spelling errors, and attackers could easily fill in the placeholder information to create a very sophisticated attack in seconds.

Unfortunately, as tools like DeepSeek become more accessible, threat actors can leverage AI to craft hyper-personalized, real-time phishing attacks at an unprecedented scale. The only effective countermeasure? AI-powered security solutions that adapt as quickly as AI-driven threats evolve.

One of the key differentiators of DeepSeek AI is its affordability. DeepSeek-V3, for instance, operates at a fraction of the cost of GPT-4o while still matching or surpassing its performance across multiple benchmarks. This lower barrier to entry means cybercriminals can leverage DeepSeek AI to generate phishing content at scale without significant investment.

Additionally, unlike OpenAI’s models, which are hosted within controlled environments with API-based access restrictions, DeepSeek AI is an open-source model. This means it can be freely downloaded, modified, and deployed on private infrastructure, eliminating oversight mechanisms designed to prevent abuse. Threat actors can therefore manipulate the model to evade security controls, making detection and mitigation significantly more challenging.

The reason why these AI-powered attacks, including those from DeepSeek AI, continue to reach end users is that legacy systems weren't set up to defend against them. The signatures that secure email gateways (SEGs) rely on to prevent attacks fail against AI-driven, text-based phishing. Further, SEGs that require constant rule updates will consistently miss adaptive attacks that are designed to bypass them.

In contrast, at Abnormal, we have designed our cyberattack detection systems to be resilient to these kinds of next-generation attacks powered by AI. Key components of the Abnormal AI platform that make it possible to detect attacks—no matter whether they’re generated by humans or AI—include:

• Behavioral AI: Understands communication patterns within an organization to identify subtle anomalies and detect never-before-seen threats in real time, even when no malicious links or attachments are present.

• Identity and Vendor Intelligence: Tracks normal behavior across employees and vendors, flagging suspicious deviations in tone, urgency, and sender history.

• Adaptive Defense Mechanisms: Ensure that Abnormal continuously updates its detection models as bad actors find new ways to utilize AI to refine their attacks.

Additionally, Abnormal’s automated response enables the platform to detect anomalies that indicate an attack, preventing end users from engaging with it—so humans never have to make a decision on whether an email is malicious or not.

The bottom line is that as attackers innovate with AI, so does Abnormal. We fight AI-powered phishing with AI-powered defense, ensuring organizations stay protected against this new era of cyber threats—whether they come from ChatGPT, Deep Seek, or humans themselves.

Want to see how Abnormal detects DeepSeek AI-powered phishing in your environment? Book a demo today and receive an AI-powered risk assessment for your organization.