These Are the Top Generative AI Dangers to Watch For in 2024

In 2023, we saw Generative AI exploitation surge, as bad actors got a host of new tools to use to create and distribute harmful content. With the lessons learned last year, here’s what we’re looking out for in 2024. Over the past year, our team, like many others, has been entrenched in GenAI. Unlike other teams, however, we’ve spent the year studying how GenAI is abused by bad actors, as we work with our partners, including seven leading foundation model organizations and GenAI applications, to ensure AI safety by design. As we’ve done so - we’ve learned a lot about bad actor tactics, foundation model loopholes, and how their convergence allows harmful content creation and distribution - at scale. With that knowledge, here are the seven generative AI risks we are preparing for in 2024:

1. Exploitation of Multimodal Capabilities

2024 will see an acceleration of AI model releases, with multimodal capabilities that allow various combinations of inputs (such as text and images in the same prompt), creating a new suite of Generative AI risks. Threat actors can combine two prompts that are benign in isolation to generate harmful materials not detected by existing safety mitigations. This is a form of adversarial machine learning that exploits how models process combined inputs to bypass individual-input safety filters. To illustrate, consider a non-violative prompt involving adult speech, combined with a childlike audio or image. Taken separately, these would not raise any flags, it is only when looking at both inputs together, that harmful content with CSAM implications may be identified. This threat extends far beyond CSAM production, for example, our testing also showed that combinations of this type can generate personal information of social media users, such as home addresses or phone numbers.

2. GenAI Audio Impersonation

Text-to-Audio models will grow more prevalent in 2024, and their use in fraud, misinformation, and other risky contexts will grow. Already being used in voice-cloning fraud schemes, this type of abuse will become more prominent in other abuse areas as well. For example, in hate speech and disinformation, threat actors may clone an individual’s voice to falsely claim that they made a controversial or misleading statement and use this to achieve malicious aims. Concerningly, Alice has also already documented voice cloning being used by child predators for malicious purposes.

3. Wide Reach of Election Disinformation

Half of the world will vote in 2024, and AI-generated mis- and disinformation will affect the electoral results. Alice's work in 2023 showed how this process is already in motion. In Q4 2023, Alice detected both foreign and domestic actors using Generative AI to target American voters with divisive narratives related to US foreign and economic policy. The potential scale for this type of abuse is dramatic: in the past month alone, the Alice team reviewed politically charged AI-generated content related to one model, with over 83M impressions.

4. Continued Abuse of GenAI Tools by Child Sexual Predators

Demonstrated early last year, the use of GenAI tools to create CSAM and sexually explicit content continues to grow. We expect this trend to continue well into next year. Additionally, changes in the sexual extortion landscape, alongside the rise in generative AI adoption, point to an upcoming explosion of nude image creation by organized crime organizations, people known to victims and other threat actors for use in extortion. This growth has already been documented by Alice in 2023, consider the following statistics:

• Dark web chatter on creating AI-generated NCII of women increased by 360%.
• GenAI CSAM producers emerged and increased by 335% across dark web predator communities.

5. IP & Copyright Infringement Liabilities

Copyright issues are a known challenge in the Generative AI space due to the fundamentals of the technology. In the context of recent lawsuits, we expect continued legal scrutiny, attempts to transfer liability across the ecosystem, and new standards and policies adopted by the major players. Generative AI players are proactively preparing for this, as in Q4, many of our AI safety customers requested deeper work in this arena.

6. Malicious LLM and Chatbot Proliferation

In 2023, we saw widespread abuse of foundation models to create bad AI models and chatbots with few to no safety restrictions (e.g., WormGPT and FraudGPT). We expect this trend to continue in 2024, as threat actors uncover more ways to exploit new— and open-source—technologies. Open-source releases lower the barrier to discovering LLM attack vectors and sharing them across dark web communities. An example of this abuse was uncovered in December 2023, when uncensored chatbots named TrumpAI, BidenAI, and BibiBot AI, which claimed to emulate politicians but actually promoted far-right and antisemitic content, were released on Gab.

7. Enterprise GenAI Application Deployment Risks

As the launch of LLM-based enterprise applications moves beyond the early adopter phase, Alice expects more incidents related to privacy, security, and Generative AI risks applicable in a corporate context. Alice customers and prospects are increasingly concerned about brand risk, the provision of problematic financial, legal, and medical advice, PII, and model jailbreaking for fraudulent purposes, and prompt injection attacks, where hidden instructions embedded in inputs redirect model behavior to serve attacker goals. Knowing what we know about AI risks in 2023, it is doubly crucial to prepare for 2024’s AI risks by proactively identifying emerging threats. Our proactive approach to AI safety includes AI red teaming and LLM red teaming across text, image, and audio modalities, risky prompt feeds, threat intelligence, prompt and output filters, and a safety management platform - enabling robust AI Safety programs and providing ongoing support and services. Click below to learn more about how Alice ensures AI safety by design.