Which practical AI threats are currently present? The answer is above zero – and they are becoming more sophisticated.
22 Apr 2025
 •Â
,
3 min. read
The inevitable has occurred – highly advanced AI technology has deviated from its intended use and is now being employed against unsuspecting targets, having navigated a grey area between morality and immorality. This embodies the paradox where well-developed technology can be repurposed for malicious intent. Here’s how these attacks are carried out.
Most of the noteworthy AI models have built-in “ethical barriers” preventing them from performing malicious actions, akin to the digital version of the Hippocratic Oath to “do no harm” first. If you were to inquire about constructing a weapon, for instance, these models have been pre-programmed to refrain from offering highly precise instructions that could facilitate significant harm.
While direct inquiries about weapon construction are prohibited, it’s still possible to enhance questioning strategies through the use of various tools to ultimately arrive at the desired information.
One effective method involves making API requests programmatically. Some recently launched initiatives focus on utilizing the backend API of an AI model to gain root access on servers. Another approach involves leveraging the backend of ChatGPT to intelligently identify potential targets for future attacks.
By combining AI-powered tools with other resources designed to address different challenges, such as bypassing obscured IPs to identify the actual target server, a potent arsenal can be created, particularly as these processes become more automated.
In the digital realm, these techniques can be harnessed to develop amalgamated tools that pinpoint vulnerabilities, iterate potential exploits, and deceive the constituent AI models effectively.
This method is somewhat analogous to a “clean room design,” where an AI model is tasked with solving a small component of a larger malicious objective defined by an attacker, and then a combination creates the final weapon.
From a legal perspective, various organizations are working to implement effective barriers that will impede these malicious activities or impose penalties on AI models that have played a role, to some extent. However, assigning precise degrees of blame is challenging. Allocating blame proportionately, particularly in the context of legal burden of proof, will be a complex undertaking.
Pioneering New Frontiers
AI models can search vast code repositories in existing software, identifying insecure patterns and devising digital weapons to exploit vulnerable software running on devices worldwide. This process presents fresh targets for compromise and provides a platform for launching zero-day attacks.
It’s conceivable that nations will intensify such endeavors – preemptively weaponizing software vulnerabilities using AI. This strategy puts defenders at a disadvantage, triggering a digital defense escalation that appears somewhat dystopian. Defenders will need to deploy their own AI-based defenses to counter these threats effectively or prevent potential breaches. Hopefully, they will rise to the challenge.
Even existing AI models can solve complex problems effortlessly, simulating human reasoning in a coherent manner (at least during clear moments). While these technologies are unlikely to evolve into sentient collaborators in the near future, their vast data consumption from the internet enables them to possess extensive knowledge that can be exploited for gathering information.
These systems will continue to enhance their capabilities, potentially reducing the need for extensive supervision. They may aid individuals devoid of moral constraints in achieving significant impact and empower resourceful actors to operate on a grand scale. Early signs of these developments have already surfaced in red team exercises or have been observed in real-world scenarios.
One thing is certain: the frequency of intelligence-driven attacks will rise. Once an exploitable CVE is disclosed or a new tactic emerges, quick thinking will be crucial – are you prepared?
About Author
