Tibor Kiss

In today’s complex digital world, building truly intelligent applications requires more than just raw data — you need to understand the intricate relationships within that data. Graph analysis helps reveal these hidden connections, and when combined with techniques like full-text search and vector search, enables you to deliver a new class of AI-enabled application experiences. However, traditional approaches based on niche tools result in data silos, operational overhead, and scalability challenges. That’s why we introduced Spanner Graph, and today we’re excited to announce that it’s generally available.

In a previous post, we described how Spanner Graph reimagines graph data management with a unified database that integrates graph, relational, search, and gen AI capabilities with virtually unlimited scalability. With Spanner Graph, you gain access to an intuitive ISO Standard Graph Query Language (GQL) interface that simplifies pattern matching and relationship traversal. You also benefit from full interoperability between GQL and SQL, for tight integration between graph and tabular data. Powerful vector and full-text search enable fast data retrieval using semantic meaning and keywords. And you can rely on Spanner’s scalability, availability, and consistency to provide a solid data foundation. Finally, integration with Vertex AI gives you access to powerful AI models directly within Spanner Graph.

What’s new in Spanner Graph

Since the preview, we have added exciting new capabilities and partner integrations to make it easier for you to build with Spanner Graph. Let’s take a closer look.

1) Spanner Graph Notebook: Graph visualization is key to developing with graphs. The new open-source Spanner Graph Notebook tool provides an efficient way to query Spanner Graph visually. This tool is natively installed in Google Colab, meaning you can use it directly within that environment. You can also leverage it in notebook environments like Jupyter Notebook. With this tool, you can use magic commands with GQL to visualize query results and graph schemas with multiple layout options, inspect node and edge properties, and analyze neighbor relationships.

2) GraphRAG with LangChain integration: Spanner Graph’s integration with LangChain allows for quick prototyping of GraphRAG applications. Conventional RAG, while capable of grounding the LLM by providing relevant context from your data using vector search, cannot leverage the implicit relationships present in your data. GraphRAG overcomes this limitation by constructing a graph from your data that captures these complex relationships. At retrieval time, GraphRAG uses the combined power of graph queries with vector search to provide a richer context to the LLM, enabling it to generate more accurate and relevant answers.

3) Graph schema in Spanner Studio: The Spanner Studio Explorer panel now displays a list of defined graphs, their nodes and edges, and the associated labels and properties. You can explore and understand the structure of your graph data at a glance, making it easier to design, debug, and optimize your applications.

4) Graph query improvements: Spanner Graph now supports the path data type and functions, allowing you to retrieve and analyze the specific sequence of nodes and relationships that connect two nodes in your graph. For example, you can create a path variable in a path pattern, using the IS_ACYCLIC function to check if the path has repeating nodes, and return the entire path:

5) Graph visualization partner integrations:  Spanner Graph is now integrated with leading graph visualization partners. For example, Spanner Graph customers can use GraphXR, Kineviz’s flagship product, which combines cutting-edge visualization technology with advanced analytics to help organizations make sense of complex, connected data. 

“We are thrilled to partner with Google Cloud to bring graph analytics to big data. By integrating GraphXR with Spanner Graph, we’re empowering businesses to visualize and interact with their data in ways that were previously unimaginable.” – Weidong Yang, CEO, Kineviz

Similarly, you can use Graphistry’s GPU-accelerated visual graph intelligence platform to extract meaningful insights from large complex data in Spanner Graph. 

“Businesses can finally handle graph data with both speed and scale. By combining Graphistry’s GPU-accelerated graph visualization and AI with Spanner Graph’s global-scale querying, teams can now easily go all the way from raw data to graph-informed action. Whether detecting fraud, analyzing journeys, hunting hackers, or surfacing risks, this partnership is enabling teams to move with confidence.” – Leo Meyerovich, Founder and CEO, Graphistry

Furthermore, you can use G.V(), a quick-to-install graph database client, with Spanner Graph to perform day-to-day graph visualization and data analytics tasks with ease. Data professionals benefit from high-performance graph visualization, no-code data exploration, and highly customizable data visualization options. 

“Graphs thrive on connections, which is why I’m so excited about this new partnership between G.V() and Google Cloud Spanner Graph. Spanner Graph turns big data into graphs, and G.V() effortlessly turns graphs into interactive data visualizations. I’m keen to see what data professionals build combining both solutions.” – Arthur Bigeard, Founder, gdotv Ltd.

What customers are saying

Through our road to GA, we have also been working with multiple customers to help them innovate with Spanner Graph:

“The Commercial Financial Network manages commercial credit data for more than 30 million U.S. businesses  Managing the hierarchy of these businesses can be complex due to the volume of these hierarchies, as well as the dynamic nature driven by mergers and acquisitions, Equifax is committed to providing lenders with the accurate, reliable and timely information they need as they make financial decisions. Spanner Graph helps us manage these rapidly changing, dynamic business hierarchies easily at scale.” – Yuvaraj Sankaran, Chief Architect of Global Platforms, Equifax

“As we strive to enhance our fraud detection capabilities, having a robust, multi-model database like Google Spanner is crucial for our success. By integrating SQL for transactional data management with advanced graph data analysis, we can efficiently manage and analyze evaluated fraud data. Spanner’s new capabilities significantly improve our ability to maintain data integrity and uncover complex fraud patterns, ensuring our systems are secure and reliable.” – Hai Sadon, Data Platform Group Manager, Transmit Security

“Spanner Graph has provided a novel and performant way for us to query this data, allowing us to deliver features faster and with greater peace of mind. Its flexible data modeling and high-performance querying have made it far easier to leverage the vast amount of data we have in our online applications.” – Aaron Tang, Senior Principal Engineer, U-NEXT 

Get started with Spanner Graph today

With Spanner Graph, we’re excited to offer graph data management alongside relational, search, and AI capabilities, all on a single unified, highly scalable database. Learn more about Spanner Graph benefits and use cases here. Use this quick setup guide to get started with Spanner Graph capabilities. You can also try out sample applications related to financial investing, fraud detection, and customer 360 and product recommendations.

Are you a cloud architect or IT admin tasked with ensuring deployments are following best practices and generating configuration validation reports? The struggle of adopting best practices is real. And not just the first time: ensuring that a config doesn’t drift from org-wide best practices over time is notoriously difficult.    

Workload Manager provides a rule-based validation service for evaluating your workloads running on Google Cloud. Workload Manager scans your workloads, including SAP and Microsoft SQL Server, to detect deviations from standards, rules, and best practices to improve system quality, reliability, and performance. .

Introducing custom rules in Workload Manager

Today, we’re excited to extend Workload Manager with custom rules (GA), a detective-based service that helps ensure your validations are not blocking any deployments, but that allows you to easily detect compliance issues across different architectural intents. Now, you can flexibly and consistently validate your Google Cloud deployments across Projects, Folders and Orgs against best practices and custom standards to help ensure that they remain compliant.

Here’s how to get started with Workload Manager custom rules in a matter of minutes.

1) Codify best practices and validate resources
Identify best practices relevant to your deployments from the Google Cloud Architecture Framework, codify them in Rego, a declarative policy language that’s used to define rules and express policies over complex data structures, and run or schedule evaluation scans across your deployments. 

Start with sample Rego policies in our GitHub repository: https://github.com/GoogleCloudPlatform/workload-manager

You can create new Rego rules based on your preferences, or reach out to your account team to get more help crafting new rules.

2) Export findings to BigQuery dataset and visualize them using Looker
You can configure your own BigQuery dataset to export each validation scan and easily integrate it with your existing reporting systems, build a new Looker dashboard, or export results to Google Sheets to plan remediation steps.  

Additionally, you can configure Pub/Sub-based notifications to send email, Google Chat messages, or integrate with your third-party systems based on different evaluation success criteria.

A flexible system to do more than typical config validation

With custom rules you can build rules with complex logic and validation requirements across multiple domains. You can delegate build and management to your subject matter experts, reducing development time and accelerating the time to release new policies. 

And with central BigQuery table export, you can combine violation findings from multiple evaluations and easily integrate with your reporting system to build a central compliance program.

Get started today with custom rules in Workload Manager by referring to the documentation and testing sample policies against your deployments.

Need more help? Engage with your account teams to get more help in crafting, curating and adopting best practices. 

Rapid advancements in artificial intelligence (AI) are unlocking new possibilities for the way we work and accelerating innovation in science, technology, and beyond. In cybersecurity, AI is poised to transform digital defense, empowering defenders and enhancing our collective security. Large language models (LLMs) open new possibilities for defenders, from sifting through complex telemetry to secure coding, vulnerability discovery, and streamlining operations. However, some of these same AI capabilities are also available to attackers, leading to understandable anxieties about the potential for AI to be misused for malicious purposes. 

Much of the current discourse around cyber threat actors’ misuse of AI is confined to theoretical research. While these studies demonstrate the potential for malicious exploitation of AI, they don’t necessarily reflect the reality of how AI is currently being used by threat actors in the wild. To bridge this gap, we are sharing a comprehensive analysis of how threat actors interacted with Google’s AI-powered assistant, Gemini. Our analysis was grounded by the expertise of Google’s Threat Intelligence Group (GTIG), which combines decades of experience tracking threat actors on the front lines and protecting Google, our users, and our customers from government-backed attackers, targeted 0-day exploits, coordinated information operations (IO), and serious cyber crime networks.

We believe the private sector, governments, educational institutions, and other stakeholders must work together to maximize AI’s benefits while also reducing the risks of abuse. At Google, we are committed to developing responsible AI guided by our principles, and we share resources and best practices to enable responsible AI development across the industry. We continuously improve our AI models to make them less susceptible to misuse, and we apply our intelligence to improve Google’s defenses and protect users from cyber threat activity. We also proactively disrupt malicious activity to protect our users and help make the internet safer. We share our findings with the security community to raise awareness and enable stronger protections for all. 

Executive Summary

Google Threat Intelligence Group (GTIG) is committed to tracking and protecting against cyber threat activity. We relentlessly defend Google, our users, and our customers by building the most complete threat picture to disrupt adversaries. As part of that effort, we investigate activity associated with threat actors to protect against malicious activity, including the misuse of generative AI or LLMs. 

This report shares our findings on government-backed threat actor use of the Gemini web application. The report encompasses new findings across advanced persistent threat (APT) and coordinated information operations (IO) actors tracked by GTIG. By using a mix of analyst review and LLM-assisted analysis, we investigated prompts by APT and IO threat actors who attempted to misuse Gemini.

GTIG takes a holistic, intelligence-driven approach to detecting and disrupting threat activity, and our understanding of government-backed threat actors and their campaigns provides the needed context to identify threat enabling activity. We use a wide variety of technical signals to track government-backed threat actors and their infrastructure, and we are able to correlate those signals with activity on our platforms to protect Google and our users. By tracking this activity, we’re able to leverage our insights to counter threats across Google platforms, including disrupting the activity of threat actors who have misused Gemini. We also actively share our insights with the public to raise awareness and enable stronger protections across the wider ecosystem.

Our analysis of government-backed threat actor use of Gemini focused on understanding how threat actors are using AI in their operations and if any of this activity represents novel or unique AI-enabled attack or abuse techniques. Our findings, which are consistent with those of our industry peers, reveal that while AI can be a useful tool for threat actors, it is not yet the game-changer it is sometimes portrayed to be. While we do see threat actors using generative AI to perform common tasks like troubleshooting, research, and content generation, we do not see indications of them developing novel capabilities. 

Our key findings include:

• We did not observe any original or persistent attempts by threat actors to use prompt attacks or other machine learning (ML)-focused threats as outlined in the Secure AI Framework (SAIF) risk taxonomy. Rather than engineering tailored prompts, threat actors used more basic measures or publicly available jailbreak prompts in unsuccessful attempts to bypass Gemini’s safety controls. 
• Threat actors are experimenting with Gemini to enable their operations, finding productivity gains but not yet developing novel capabilities. At present, they primarily use AI for research, troubleshooting code, and creating and localizing content. 
• APT actors used Gemini to support several phases of the attack lifecycle, including researching potential infrastructure and free hosting providers, reconnaissance on target organizations, research into vulnerabilities, payload development, and assistance with malicious scripting and evasion techniques. Iranian APT actors were the heaviest users of Gemini, using it for a wide range of purposes. Of note, we observed limited use of Gemini by Russian APT actors during the period of analysis.
• IO actors used Gemini for research; content generation including developing personas and messaging; translation and localization; and to find ways to increase their reach. Again, Iranian IO actors were the heaviest users of Gemini, accounting for three quarters of all use by IO actors. We also observed Chinese and Russian IO actors using Gemini primarily for general research and content creation. 
• Gemini’s safety and security measures restricted content that would enhance adversary capabilities as observed in this dataset. Gemini provided assistance with common tasks like creating content, summarizing, explaining complex concepts, and even simple coding tasks. Assisting with more elaborate or explicitly malicious tasks generated safety responses from Gemini. 
• Threat actors attempted unsuccessfully to use Gemini to enable abuse of Google products, including researching techniques for Gmail phishing, stealing data, coding a Chrome infostealer, and bypassing Google’s account verification methods. 

Rather than enabling disruptive change, generative AI allows threat actors to move faster and at higher volume. For skilled actors, generative AI tools provide a helpful framework, similar to the use of Metasploit or Cobalt Strike in cyber threat activity. For less skilled actors, they also provide a learning and productivity tool, enabling them to more quickly develop tools and incorporate existing techniques. However, current LLMs on their own are unlikely to enable breakthrough capabilities for threat actors. We note that the AI landscape is in constant flux, with new AI models and agentic systems emerging daily. As this evolution unfolds, GTIG anticipates the threat landscape to evolve in stride as threat actors adopt new AI technologies in their operations.

AI-Focused Threats

Attackers can use LLMs in two ways. One way is attempting to leverage LLMs to accelerate their campaigns (e.g., by generating code for malware or content for phishing emails). The overwhelming majority of activity we observed falls into this category. The second way attackers can use LLMs is to instruct a model or AI agent to take a malicious action (e.g., finding sensitive user data and exfiltrating it). These risks are outlined in Google’s Secure AI Framework (SAIF) risk taxonomy. 

We did not observe any original or persistent attempts by threat actors to use prompt attacks or other AI-specific threats. Rather than engineering tailored prompts, threat actors used more basic measures, such as rephrasing a prompt or sending the same prompt multiple times. These attempts were unsuccessful.

Jailbreak Attempts: Basic and Based on Publicly Available Prompts

We observed a handful of cases of low-effort experimentation using publicly available jailbreak prompts in unsuccessful attempts to bypass Gemini’s safety controls. Threat actors copied and pasted publicly available prompts and appended small variations in the final instruction (e.g., basic instructions to create ransomware or malware). Gemini responded with safety fallback responses and declined to follow the threat actor’s instructions. 

In one example of a failed jailbreak attempt, an APT actor copied publicly available prompts into Gemini and appended basic instructions to perform coding tasks. These tasks included encoding text from a file and writing it to an executable and writing Python code for a distributed denial-of-service (DDoS) tool. In the former case, Gemini provided Python code to convert Base64 to hex, but provided a safety filtered response when the user entered a follow-up prompt that requested the same code as a VBScript.

The same group used a different publicly available jailbreak prompt to request Python code for DDoS. Gemini provided a safety filtered response stating that it could not assist, and the threat actor abandoned the session and did not attempt further interaction.

Some malicious actors unsuccessfully attempted to prompt Gemini for guidance on abusing Google products, such as advanced phishing techniques for Gmail, assistance coding a Chrome infostealer, and methods to bypass Google’s account creation verification methods. These attempts were unsuccessful. Gemini did not produce malware or other content that could plausibly be used in a successful malicious campaign. Instead, the responses consisted of safety-guided content and generally helpful, neutral advice about coding and cybersecurity. In our continuous work to protect Google and our users, we have not seen threat actors either expand their capabilities or better succeed in their efforts to bypass Google’s defenses.

Findings: Government-Backed Threat Actors Misusing Gemini

Google analyzed Gemini activity associated with known APT actors and identified APT groups from more than 20 countries that used Gemini. The highest volume of usage was from Iran and China. APT actors used Gemini to support several phases of the attack lifecycle, including researching potential infrastructure and free hosting providers, reconnaissance on target organizations, research into vulnerabilities, payload development, and assistance with malicious scripting and evasion techniques. The top use cases by APT actors focused on: 

• Assistance with coding tasks, including troubleshooting, tool and script development and converting or rewriting existing code 

• Vulnerability research focused on publicly reported vulnerabilities and specific CVEs 

• General research on various technologies, translations and technical explanations 

• Reconnaissance about likely targets, including details about specific organizations 

• Enabling post-compromise activity, such seeking advice on techniques to evade detection, escalate privileges or conduct internal reconnaissance in a target environment

We observed APT actors use Gemini attempting to support all phases of the attack lifecycle.

Iranian Government-Backed Actors 

Iranian government-backed actors accounted for the largest Gemini use linked to APT actors. Across Iranian government-backed actors, we observed a broad scope of research and use cases, including to enable reconnaissance on targets, for research into publicly reported vulnerabilities, to request translation and technical explanations, and to create content for possible use in future campaigns. Their use reflected strategic Iranian interests including research focused on defense organizations and experts, defense systems, foreign governments, individual dissidents, the Israel-Hamas conflict, and social issues in Iran.

Crafting Phishing Campaigns 

Over 30% of Iranian APT actors’ Gemini use was linked to APT42. APT42’s Gemini activity reflected the group’s focus on crafting successful phishing campaigns. We observed the group using Gemini to conduct reconnaissance into individual policy and defense experts, as well as organizations of interest for the group. 

In addition to reconnaissance, APT42 used the text generation and editing capabilities of Gemini to craft material for phishing campaigns, including generating content with cybersecurity themes and tailoring the output to a US defense organization. APT42 also utilized Gemini for translation including localization, or tailoring content for a local audience. This includes content tailored to local culture and local language, such as asking for translations to be in fluent English.

Vulnerability Research

The majority of APT42’s efforts focused on research into publicly known vulnerabilities, such as a request to generate a list of critical vulnerabilities from 2023. They also focused on vulnerabilities in specific products such as Mikrotik, Apereo, and Atlassian.

Of note, APT42 appeared to be researching how to use generative AI tools for offensive purposes, asking Gemini for help preparing training content for a red team focused on how offensive teams can use AI tools in their operations.  

Research Into Military and Weapons Systems 

APT42 also appears to have used Gemini’s translation and explanation functions to better understand publicly available information on defense systems. Their efforts included general research into the Israel-Hamas conflict, as well as strategic trends in China’s defense industry. The threat actor also used Gemini for technical explanations about US-made aerospace systems.

Another Iranian APT group also focused on understanding warfare defenses including specific research into satellite signal jamming and anti-drone systems. Other Iranian APT actors researched specific defense systems, including researching information about specific unmanned aerial vehicle (UAV) models, jamming F-35 fighter jets, anti-drone systems, and Israel’s missile defense systems. 

People’s Republic of China (PRC) Government-Backed Actors

Government-backed actors linked to the People’s Republic of China (PRC) attempted to use Gemini to enable reconnaissance on targets, for scripting and development, to request translation and explanation of technical concepts, and attempting to enable deeper access to a network following initial compromise. PRC threat actors’ usage resembled an IT admin seeking to streamline, troubleshoot, or automate their tasks. In a malicious context, however, this activity could be used to enable lateral movement, privilege escalation, data exfiltration, and detection evasion.

Enabling Deeper Access in a Target Network 

PRC-backed APT actors also used Gemini to work through scripting and development tasks, many of which appeared intended to enable deeper access in a target network after threat actors obtained initial access. For example, one PRC-backed group asked Gemini for assistance figuring out how to sign a plugin for Microsoft Outlook and silently deploy it to all computers. The same actor also asked Gemini to generate code to remotely access Windows Event Log; sought instructions on how to add a self-signed certificate to Active Directory; and asked Gemini for a command to identify the IP addresses of administrators on the domain controller. Other actors used Gemini for help troubleshooting Chinese character encoding issues in smbclient and how to record passwords on the VMware vCenter. 

In another example, PRC-backed APT actors asked Gemini for assistance with Active Directory management commands and requested help troubleshooting impacket, a Python-based tool for working with network protocols. While impacket is commonly used for benign purposes, the context of the threat actor made it clear that the actor was using the tool for malicious purposes. 

Explaining Tools, Concepts, and Code 

PRC actors utilized Gemini to learn about specific tools and technologies and develop solutions to technical challenges. For example, a PRC APT actor used Gemini to break down how to use the graph database Nebula Graph. In another instance, the same actor used Gemini to offer possible solutions to TLS 1.3 visibility challenges. Another PRC-backed APT group sought to understand a malicious PHP script. 

Vulnerability Research and Reverse Engineering

In one case, a PRC-backed APT actor attempted unsuccessfully to get Gemini’s help reverse engineering the endpoint detection and response (EDR) tool Carbon Black. The same threat actor copied disassembled Python bytecode into Gemini to convert the bytecode into Python code. It’s not clear what their objective was. 

Unsuccessful Attempts to Elicit Internal System Information From Gemini

In one case, the PRC-backed APT actor APT41 attempted unsuccessfully to use Gemini to learn about Gemini’s underlying infrastructure and systems. The actor asked Gemini to share details such as its IP address, kernel version, and network configuration. Gemini responded but did not disclose sensitive information. In a helpful tone, the responses provided publicly available details that would be widely known about the topic, while also indicating that the requested information is kept secret to prevent unauthorized access. 

North Korean Government-Backed Actors

North Korean APT actors used Gemini to support several phases of the attack lifecycle, including researching potential infrastructure and free hosting providers, reconnaissance on target organizations, payload development, and assistance with malicious scripting and evasion techniques. They also used Gemini to research topics of strategic interest to the North Korean government, such as South Korean nuclear technology and cryptocurrency. We also observed that North Korean actors were using LLMs in likely attempts to enable North Korea’s efforts to place clandestine IT workers at Western companies.

Clandestine IT Worker Threat

North Korean APT actors used Gemini to draft cover letters and research jobs—activities that would likely support efforts by North Korean nationals to use fake identities and obtain freelance and full-time jobs at foreign companies while concealing their true identities and locations. One North Korea-backed group utilized Gemini to draft cover letters and proposals for job descriptions, researched average salaries for specific jobs, and asked about jobs on LinkedIn. The group also used Gemini for information about overseas employee exchanges. Many of the topics would be common for anyone researching and applying for jobs. 

While normally employment-related research would be typical for any job seeker, we assess the usage is likely related to North Korea’s ongoing efforts to place clandestine workers in freelance gigs or full-time jobs at Western firms. The scheme, which involves thousands of North Korean workers and has affected hundreds of US-based companies, uses IT workers with false identities to complete freelance work and send wages back to the North Korean regime.

Target Research and Reconnaissance

North Korean actors also engaged with Gemini with several questions that appeared focused on conducting initial research and reconnaissance into prospective targets. They also researched organizations and industries that are typical targets for North Korean actors, including the US and South Korean militaries and defense contractors. One North Korean APT group asked Gemini for information about companies and organizations across a variety of industry sectors and regions. Some of this Gemini usage related directly to organizations that the same group had attempted to target in phishing and malware campaigns that Google previously detected and disrupted.

In addition to research into companies, North Korean APT actors researched nuclear technology and power plants in South Korea, such as site locations, recent news articles, and the security status of the plants. Gemini responded with widely available, public information and facts that would be easily discoverable in an online search. 

Help with Scripting, Payload Development, Defense Evasion 

North Korean actors also tried to use Gemini to assist with development and scripting tasks. One North Korea-backed group attempted to use Gemini to help develop webcam recording code in C++. Gemini provided multiple versions of code, and repeated efforts by the actor potentially suggested their frustration by Gemini’s answers. The same group also asked Gemini to generate a robots.txt file to block crawlers and an .htaccess file to redirect all URLs except CSS extensions. 

One North Korean APT actor used Gemini for assistance developing code for sandbox evasion. For example, the threat actor utilized Gemini to write code in C++ to detect VM environments and Hyper-V virtual machines. Gemini provided responses with short code snippets to perform simple sandbox checks. The same group also sought help troubleshooting Java errors when implementing AES encryption, and separately asked Gemini if it is possible to acquire a system password on Windows 11 using Mimikatz. 

Russian Government-Backed Actors 

During the period of analysis, we observed limited use of Gemini by Russia-backed APT actors. Of this limited use, the majority of usage appeared benign, rather than threat-enabling. The reasons for this low engagement are unclear. It is possible Russian actors avoided Gemini out of operational security considerations, staying off Western-controlled platforms to avoid monitoring of their activities. They may be using AI tools produced by Russian firms or locally hosting LLMs, which would ensure full control of their infrastructure. Alternatively, they may have favored other Western LLMs.

One Russian government-backed group used Gemini to request help with a handful of tasks, including help rewriting publicly available malware into another language, adding encryption functionality to code, and explanations for how a specific block of publicly available malicious code functions.

Financially Motivated Actors Using LLMs

Threat actors in underground marketplaces are advertising ways to bypass security guardrails to help LLMs with malware development, phishing, and other malicious tasks. The offerings include jailbroken LLMs that are ready-made for malicious use. 

Throughout 2023 and 2024, Google Threat Intelligence Group (GTIG) observed underground forum posts related to LLMs, indicating there is a burgeoning market for nefarious versions of LLMs. Some advertisements boast customized and jailbroken LLMs that don’t have restrictions for malware development purposes, or they tout a lack of security measures typically found on legitimate services, allowing the user to prompt the LLM about any topic or task without incurring security guardrails or limits on their queries. Examples include FraudGPT, which has been advertised on Telegram as having no limitations, and WormGPT, a privacy focused, “uncensored” LLM capable of developing malware.   

Financially motivated actors are using LLMs to help augment business email compromise (BEC) operations. GTIG has noted evidence of financially motivated actors using manipulated video and voice content in business email compromise (BEC) scams. Media reports indicate that financially motivated actors have reportedly used WormGPT to create more persuasive BEC messages.

Findings: Information Operations (IO) Actors Misusing Gemini

IO actors used Gemini for research, content generation including developing personas and messaging, translation and localization, and to find ways to increase their reach. Common use cases include general research into news and current events as well as specific research into individuals and organizations. In addition to creating content for campaigns, including personas and content, the actors researched increasing the efficacy of campaigns, including automating distribution, using search engine optimization (SEO) to optimize the reach of campaigns, and increasing operational security. As with government-backed groups, IO actors also used Gemini for translation and localization and for understanding the meanings or context of content. 

Iran-Linked Information Operations Actors

Iran-based information operations (IO) groups used Gemini for a wide range of tasks, including general research, translation and localization, content creation and manipulation, and generating content with a specific bias or tone. We also observed Iran-based IO actors engage with Gemini about news events and ask Gemini to provide details on economic and political issues in Iran, the US, the Middle East, and Europe.

In line with their practice of mixing original and borrowed content, Iranian IO actors translated existing material, including news-like articles. They then used Gemini to explain the context and meaning of particular phrases within the given text.

Iran-based IO actors also used Gemini to localize the content, seeking human-like translation and asking Gemini for help with tasks like making the text sound like a native English speaker. They used Gemini to manipulate text (e.g., asking for help rewriting existing text on immigration and crime in a specific style or tone). 

Iran’s activity also included research about improving the reach of their campaigns. For example, they attempted to generate SEO-optimized content, likely in an effort to reach a larger audience. Some actors also used Gemini to suggest strategies for increasing engagement on social media.

PRC-Linked Information Operations Actors

IO actors linked to the People’s Republic of China (PRC) used Gemini primarily for general research on a wide variety of topics. The most prolific IO actor we track, the pro-China group DRAGONBRIDGE, was responsible for approximately three quarters of this activity. Of their activity, the majority use was general research about a wide variety of topics, ranging from details about the features of various social media platforms to questions about various topics of strategic interest to the PRC government. Actors researched information on current events and politics in other regions, with a focus on the US and Taiwan. They also showed interest in assessing the impact and risk of certain events. In a handful of cases, DRAGONBRIDGE used Gemini to generate articles or content on specific topics.

DRAGONBRIDGE has experimented with other generative AI tools to create synthetic content in support of their IO campaigns. As early as 2022, the group used a commercial AI service in videos on YouTube to depict AI-generated news presenters. Their use of AI-generated video continued through 2024 but has not resulted in significantly higher engagement from real viewers. Google detected and terminated the channels distributing this content immediately upon discovery. DRAGONBRIDGE’s use of AI-generated videos or images has not resulted in significantly higher engagement from real viewers.

Russia-Linked Information Operations Actors

Russian IO actors used Gemini for general research, content creation, and translation. Half of this activity was associated with the Russian IO actor we track as KRYMSKYBRIDGE, which is linked to a Russian consulting firm that works with the Russian government. Approximately 40% of activity was linked to actors associated with Russian state sponsored entities formerly controlled by the late Russian oligarch Yevgeny Prigozhin. We also observed usage by actors tracked publicly as Doppelganger. 

The majority of Russian IO actor usage was related to general research tasks, ranging from the Russia-Ukraine war to details about various tools and online services. Russian IO actors also used Gemini for content creation, rewriting article titles and planning social media campaigns. Translation to and from Russian was also a common task. 

Russian IO actors focused on the generative AI landscape, which may indicate an interest in developing native capabilities in AI on infrastructure they control. They researched tools that can be used to create an online AI chatbot and developer tools for interacting with LLMs. One Russian IO actor used Gemini to suggest options for textual content analysis. 

Pro-Russia IO actors have used AI in their influence campaigns in the past. In 2024, the actor known as CopyCop likely used LLMs to generate content, and some stories on their sites included metadata indicating an LLM was prompted to rewrite articles from genuine news sources with a particular political perspective or tone. CopyCop’s inauthentic news sites pose as US- and Europe-based news outlets and post Kremlin-aligned views on Western policy, the war in Ukraine, and domestic politics in the US and Europe.

Building AI Safely and Responsibly 

We believe our approach to AI must be both bold and responsible. To us, that means developing AI in a way that maximizes the positive benefits to society while addressing the challenges. Guided by our AI Principles, Google designs AI systems with robust security measures and strong safety guardrails, and we continuously test the security and safety of our models to improve them. Our policy guidelines and prohibited use policies prioritize safety and responsible use of Google’s generative AI tools. Google’s policy development process includes identifying emerging trends, thinking end-to-end, and designing for safety. We continuously enhance safeguards in our products to offer scaled protections to users across the globe.  

At Google, we leverage threat intelligence to disrupt adversary operations. We investigate abuse of our products, services, users and platforms, including malicious cyber activities by government-backed threat actors, and work with law enforcement when appropriate. Moreover, our learnings from countering malicious activities are fed back into our product development to improve safety and security for our AI models. Google DeepMind also develops threat models for generative AI to identify potential vulnerabilities, and creates new evaluation and training techniques to address misuse caused by them. In conjunction with this research, DeepMind has shared how they’re actively deploying defenses within AI systems along with measurement and monitoring tools, one of which is a robust evaluation framework used to automatically red team an AI system’s vulnerability to indirect prompt injection attacks. Our AI development and Trust & Safety teams also work closely with our threat intelligence, security, and modelling teams to stem misuse.

The potential of AI, especially generative AI, is immense. As innovation moves forward, the industry needs security standards for building and deploying AI responsibly. That’s why we introduced the Secure AI Framework (SAIF), a conceptual framework to secure AI systems. We’ve shared a comprehensive toolkit for developers with resources and guidance for designing, building, and evaluating AI models responsibly. We’ve also shared best practices for implementing safeguards, evaluating model safety, and red teaming to test and secure AI systems. 

About the Authors

Google Threat Intelligence Group brings together the Mandiant Intelligence and Threat Analysis Group (TAG) teams, and focuses on identifying, analyzing, mitigating, and eliminating entire classes of cyber threats against Alphabet, our users, and our customers. Our work includes countering threats from government-backed attackers, targeted 0-day exploits, coordinated information operations (IO), and serious cyber crime networks. We apply our intelligence to improve Google’s defenses and protect our users and customers.