Security Risks and Privacy Challenges of Large Language Models in AI Systems
Large language models (LLMs) present a dual-use dilemma in cybersecurity, as their capabilities can be leveraged for both defensive and offensive purposes. Security researchers have identified purpose-built malicious LLMs, such as WormGPT and KawaiiGPT, which are designed to facilitate cybercrime by generating convincing phishing content and rapidly producing or modifying malicious code. The thin line between beneficial and harmful use of LLMs is defined largely by developer intent and the presence or absence of ethical safeguards, raising concerns about the proliferation of offensive AI tools in the threat landscape.
In addition to malicious use, LLMs face significant challenges in maintaining privacy and security due to contextual integrity failures and regulatory-driven censorship. Research from Microsoft highlights the need for AI agents to respect contextual privacy norms, as current models may inadvertently leak sensitive information. Meanwhile, the DeepSeek-R1 model demonstrates how geopolitical censorship mechanisms can introduce security flaws, such as insecure code generation and broken authentication, especially when handling politically sensitive prompts. These issues underscore the urgent need for robust privacy controls and security-aware development practices in the deployment of LLM-powered systems.
Get ahead of threats like this
Mallory correlates global threat intelligence with your attack surface — know if you’re exposed before adversaries strike.
How this story unfolded
4 events from the most recent confirmed update back to the earliest known activity.
Unit 42 publishes analysis of malicious LLMs' dual-use risks
Palo Alto Networks Unit 42 published research detailing how malicious LLMs such as WormGPT 4 and KawaiiGPT can support phishing, malware scaffolding, reconnaissance, and ransomware workflows. The report argues that commercialization and democratization of these tools are making AI-enabled cybercrime more scalable.
Commercialized 'WormGPT 4' appears on Telegram and underground forums
After the original WormGPT's emergence, a successor branded 'WormGPT 4' was promoted as a subscription service through Telegram and underground forums. Unit 42 describes this as a sign of commercialization of malicious LLMs and broader access for less-skilled threat actors.
KawaiiGPT 2.5 is identified as a free GitHub-available malicious tool
Unit 42 reports that KawaiiGPT version 2.5 was identified in July 2025 as a freely available tool on GitHub. The model was described as capable of generating spear-phishing lures, lateral-movement scripts, and data-exfiltration code.
WormGPT emerges as a malicious LLM offering offensive capabilities
Unit 42 says WormGPT first emerged in July 2023 as a purpose-built malicious large language model, reportedly based on GPT-J 6B and trained on malicious datasets. It was marketed for criminal use cases such as phishing, malware development, and other offensive tasks with safety guardrails removed.
Related entities
Vulnerabilities, threat actors, malware, products, organizations, and breaches Mallory has linked to this story.
Sources
3 references tracked. Mallory keeps watching after this page renders.
The Dual-Use Dilemma of AI: Malicious LLMs
unit42.paloaltonetworks.com
Open sourceReducing Privacy leaks in AI: Two approaches to contextual integrity
microsoft.com
Open sourceDeepSeek LLM’s geopolitical censorship triggers security flaws
scworld.com
Open sourceSee the full picture, correlated to your attack surface.
Map indicators from this story to your assets and identify affected systems in minutes.
Every observed campaign, victim, and pivot linked to actors named in this story.
Malware, exploits, and IOCs connected to the activity described here.
YARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Get matching new stories delivered to your team as they break — not the next morning.
Ask questions about this story and take action on the answers.
Related stories
Other stories Mallory is tracking that overlap on entities, sources, or topic labels.
Enterprise Security Risks and Criminal Abuse of Large Language Models
The widespread integration of large language models (LLMs) into enterprise environments is introducing new security risks at every layer of the technology stack. Security leaders are being urged to rethink traditional trust boundaries, as LLMs can alter assumptions about data handling, application behavior, and internal controls. Key risks include prompt injection, sensitive data leakage through inputs and outputs, and fragmented ownership of LLM-related security responsibilities. Experts emphasize the need to treat LLMs as untrusted compute and to enforce explicit policy and validation layers, rather than relying solely on prompt engineering or fine-tuning. Meanwhile, cybercriminals are actively exploiting the popularity of LLMs by selling discounted access to mainstream AI tools such as ChatGPT, Perplexity, and Gemini on underground forums. These tools are being used by threat actors for a range of malicious activities, including phishing, reconnaissance, and automating cybercrime operations. The criminal use of LLMs lowers the barrier to entry for less-skilled attackers and enables more efficient execution of threat campaigns, highlighting the dual challenge of securing enterprise LLM deployments while monitoring their abuse in the cybercriminal ecosystem.
Mar 21, 2026
Risks of Over-Reliance and Human Factors in Large Language Model Security
The widespread adoption of large language models (LLMs) in enterprise environments has introduced significant security challenges, particularly due to the tendency to over-rely on their outputs and the normalization of risky behaviors. Experts warn that treating LLMs as reliable and deterministic can lead to systemic vulnerabilities, as these models are inherently probabilistic and can be manipulated through techniques such as indirect prompt injection. This normalization of deviance—where unsafe practices become accepted due to a lack of immediate negative consequences—mirrors historical safety failures in other industries and is exacerbated when vendors make insecure design decisions by default. In addition to technical risks, human factors play a critical role in LLM security. Employees may inadvertently expose sensitive data by pasting it into public LLMs, blindly trust AI-generated outputs, or bypass security policies for convenience, making internal misuse a primary concern. While technical controls such as AI governance and access restrictions are important, organizations must also prioritize security awareness training to address the human side of LLM risk. Building a culture of responsible AI use is essential to mitigate both external threats and internal errors associated with LLM deployment.
Mar 21, 2026
Privacy Risks and Data Handling in Large Language Models and ChatGPT
Recent research and analysis have highlighted significant concerns regarding the privacy risks associated with large language models (LLMs) and popular AI applications such as ChatGPT. A comprehensive study conducted by researchers from Carnegie Mellon University and Northeastern University reviewed over 1,300 academic papers on AI and machine learning privacy published between 2016 and 2025. The study found that the vast majority of research has focused narrowly on training data leakage and direct chat exposure, with 92 percent of papers addressing only these two areas. This leaves a critical gap in understanding and mitigating more subtle privacy risks, such as inference attacks, context leakage through LLM agents, and the aggregation of user data at scale. The researchers argue that privacy risks extend throughout the entire lifecycle of LLMs, from data collection and processing to deployment and user interaction. They also point out systemic barriers in the research community, including a lag between technological advances and policy development, and a cultural bias that undervalues privacy concerns involving human factors. Meanwhile, practical guidance for end users is becoming increasingly important as AI applications like ChatGPT collect a wide range of personal and technical information, including account details, device data, and potentially sensitive user inputs. OpenAI, the developer of ChatGPT, maintains region-specific privacy policies, but both versions allow for extensive data collection by default. Users can take steps to limit the use of their data for model training, such as configuring privacy settings and enabling features like Temporary Chats. There are also risks associated with connecting third-party services to ChatGPT, which can increase the exposure of personal data. Users are advised to manage AI memory, disable unnecessary integrations, and secure their accounts to reduce the risk of unauthorized access. The lack of cross-disciplinary collaboration between AI, policy, and human-computer interaction fields further exacerbates the challenge of addressing these privacy risks. As LLMs become more integrated into daily life and business operations, organizations and individuals must be proactive in understanding and mitigating the full spectrum of privacy threats. The research underscores the need for a broader approach to privacy that goes beyond technical fixes and includes policy, design, and user education. Both academic and practical perspectives agree that current privacy protections are insufficient for the evolving landscape of AI-driven applications. The findings call for increased attention to underexplored areas of privacy risk, more robust privacy controls for end users, and greater alignment between technological development and regulatory frameworks. Ultimately, safeguarding privacy in the age of LLMs requires a holistic strategy that addresses both technical and human factors across the entire AI ecosystem.
Mar 21, 2026