Large Loss of Money? Choose Your LLM Security Solution Wisely.

Your company has likely started using LLMs; you may be using them yourself. It’s possible other executive leadership has put a focus on building your organization’s own proprietary LLM for your customers and/or employees. Although the “first-day gains” may show efficiency and productivity improvements short term, LLMs need to be considered with a long-term lens. Integrating LLMs into your business without a thorough cybersecurity evaluation can lead to massive financial losses — the kind of losses we typically associate with breaches.

In addition, ask yourself what the reputational fallout would be if LLMs generated harmful or offensive content that affected your customer trust and brand integrity. We wrote this blog post because you’re not the only one with questions about navigating choppy AI waters.

LLMs are a complex new attack surface that most organizations don’t fully understand, and that is to be expected with any major technological transformation. We need to be thinking of our attack surface from the inside out. From prompt injections to data exfiltration, LLMs open up new vectors for sophisticated cybercriminals.

A security-focused audit of these new technologies is crucial. Performing due diligence is a good rule of thumb with any new technology, but it is particularly important with LLMs. Defenders have only scratched the surface on these techno-powerhouses, and attackers’ freedom from bureaucracy lends them an unfortunate leg up. Attackers don’t have shareholders to appease — if an attack fails, they just move on and try it a new way. You, however, need to protect your organization the first time.

We live in bits and bytes here in the cyber world, but dollar signs shouldn’t be ignored. A cyber framework is one tool among a series of tools — and, to do the job well, you need to use the correct tool. 

Most cybersecurity evaluation frameworks aren’t built to rank financial losses from AI-related attacks. These technical frameworks are just that — frameworks. They’re not encompassing. Even if they could estimate the cost of breach, they certainly wouldn’t show you how a seemingly minor vulnerability could escalate into a multimillion-dollar crisis. These risks will continue to increase as more LLMs become incorporated into our daily lives.

The time it takes for an attacker to make a massive impact is severely shortened with LLMs. Imagine an attacker manipulates your LLM into revealing proprietary data or launching harmful commands. In mere minutes, your organization could be facing financial and reputational damage on an unprecedented scale.

When an incident involving LLMs goes public, it’s not just about legal fines or regulatory action. It’s about losing customer confidence, even in the absence of a direct breach or technical failure.

Unlike traditional cyberthreats, the chat-based interface of LLMs itself poses a serious risk since the output — seemingly harmless human language — can become dangerously toxic. We are aware of how words can hurt people, but these words can hurt your technology, too.

LLMs pose a threat that is far more insidious and unpredictable than the conventional cyber risks that your security teams are equipped to handle. In traditional threats, the attack surface is usually highly technical, defined by the target's technology or specific application flows.

However, with LLMs, the "protocol" is natural language, which vastly expands the attack surface and enables manipulation techniques that go beyond classic technological vulnerabilities and venture into human-like manipulations. Social engineering has been a favored method of attack on humans for years; now, those same concepts can be used against our technology.

In our research on prompt injection vulnerabilities across multiple data sets, a large number of cases lead to reputational damage. These involve toxic, offensive, biased, or manipulated outputs, as well as unintended responses from the LLM that deviate from the LLM developer's true intent. We have all seen the negative impact toxic behavior can have on human beings, and the closer our technology mirrors our brains, the more aware we have to be of these “softer” technological threats.

LLMs are not just another tool in your tech stack — they are a wide new attack surface with unique risks. If you’re not carefully evaluating these systems for both security vulnerabilities and financial risk, you’re setting yourself up for a potential disaster. It’s essential to ensure that any LLM security solution you deploy incorporates four key aspects:

1. Prompt security controls (prompt injection protection)
2. Sandboxing and segmentation
3. Behavioral analysis integration
4. Rigorous audit trails

Your LLM security solution should include comprehensive prompt content filtering on both input and output to detect and block attempts to extract proprietary data, along with defensive prompt engineering techniques that prevent LLM manipulation. This should include contextual prompt restrictions and validation checks to safeguard against harmful command execution.

It should ensure that LLM interactions are sandboxed and that the LLM environment is segmented from critical operational systems to prevent unauthorized command executions. Microsegmentation is a key part of any Zero Trust framework, and there’s no better time than now to start moving toward Zero Trust if you haven’t already. The ability to quickly identify where lateral movement is occurring — and stop it — can be the difference between an alert and an incident.

Your LLM security product should use behavioral analytics to establish a baseline of typical LLM interactions. It should flag and verify atypical behavior to identify potentially fraudulent or unauthorized activities.

To secure LLMs, you need to think of them as humans with ridiculously fast processing power. Lean on your social engineers and “human hackers” for strategies to protect these supercharged techno-humans.

Maintaining comprehensive audit logs of LLM interactions to track inputs, outputs, and transactions is critical to successful LLM security. This can also be valuable in postincident investigations to assess and address fraudulent actions quickly. As you well know: It’s not if, it’s when you’ll be attacked, so assuming breach is a strong feather in an LLM security solution’s cap.