Blog
/
AI
/
October 30, 2023

Exploring AI Threats: Package Hallucination Attacks

Learn how malicious actors exploit errors in generative AI tools to launch packet attacks. Read how Darktrace products detect and prevent these threats!
Inside the SOC
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Written by
Charlotte Thompson
Cyber Analyst
Written by
Tiana Kelly
Deputy Team Lead, London & Cyber Analyst
Default blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog image
30
Oct 2023

AI tools open doors for threat actors

On November 30, 2022, the free conversational language generation model ChatGPT was launched by OpenAI, an artificial intelligence (AI) research and development company. The launch of ChatGPT was the culmination of development ongoing since 2018 and represented the latest innovation in the ongoing generative AI boom and made the use of generative AI tools accessible to the general population for the first time.

ChatGPT is estimated to currently have at least 100 million users, and in August 2023 the site reached 1.43 billion visits [1]. Darktrace data indicated that, as of March 2023, 74% of active customer environments have employees using generative AI tools in the workplace [2].

However, with new tools come new opportunities for threat actors to exploit and use them maliciously, expanding their arsenal.

Much consideration has been given to mitigating the impacts of the increased linguistic complexity in social engineering and phishing attacks resulting from generative AI tool use, with Darktrace observing a 135% increase in ‘novel social engineering attacks’ across thousands of active Darktrace/Email™ customers from January to February 2023, corresponding with the widespread adoption of ChatGPT and its peers [3].

Less overall consideration, however, has been given to impacts stemming from errors intrinsic to generative AI tools. One of these errors is AI hallucinations.

What is an AI hallucination?

AI “hallucination” is a term which refers to the predictive elements of generative AI and LLMs’ AI model gives an unexpected or factually incorrect response which does not align with its machine learning training data [4]. This differs from regular and intended behavior for an AI model, which should provide a response based on the data it was trained upon.  

Why are AI hallucinations a problem?

Despite the term indicating it might be a rare phenomenon, hallucinations are far more likely than accurate or factual results as the AI models used in LLMs are merely predictive and focus on the most probable text or outcome, rather than factual accuracy.

Given the widespread use of generative AI tools in the workplace employees are becoming significantly more likely to encounter an AI hallucination. Furthermore, if these fabricated hallucination responses are taken at face value, they could cause significant issues for an organization.

Use of generative AI in software development

Software developers may use generative AI for recommendations on how to optimize their scripts or code, or to find packages to import into their code for various uses. Software developers may ask LLMs for recommendations on specific pieces of code or how to solve a specific problem, which will likely lead to a third-party package. It is possible that packages recommended by generative AI tools could represent AI hallucinations and the packages may not have been published, or, more accurately, the packages may not have been published prior to the date at which the training data for the model halts. If these hallucinations result in common suggestions of a non-existent package, and the developer copies the code snippet wholesale, this may leave the exchanges vulnerable to attack.

Research conducted by Vulcan revealed the prevalence of AI hallucinations when ChatGPT is asked questions related to coding. After sourcing a sample of commonly asked coding questions from Stack Overflow, a question-and-answer website for programmers, researchers queried ChatGPT (in the context of Node.js and Python) and reviewed its responses. In 20% of the responses provided by ChatGPT pertaining to Node.js at least one un-published package was included, whilst the figure sat at around 35% for Python [4].

Hallucinations can be unpredictable, but would-be attackers are able to find packages to create by asking generative AI tools generic questions and checking whether the suggested packages exist already. As such, attacks using this vector are unlikely to target specific organizations, instead posing more of a widespread threat to users of generative AI tools.

Malicious packages as attack vectors

Although AI hallucinations can be unpredictable, and responses given by generative AI tools may not always be consistent, malicious actors are able to discover AI hallucinations by adopting the approach used by Vulcan. This allows hallucinated packages to be used as attack vectors. Once a malicious actor has discovered a hallucination of an un-published package, they are able to create a package with the same name and include a malicious payload, before publishing it. This is known as a malicious package.

Malicious packages could also be recommended by generative AI tools in the form of pre-existing packages. A user may be recommended a package that had previously been confirmed to contain malicious content, or a package that is no longer maintained and, therefore, is more vulnerable to hijack by malicious actors.

In such scenarios it is not necessary to manipulate the training data (data poisoning) to achieve the desired outcome for the malicious actor, thus a complex and time-consuming attack phase can easily be bypassed.

An unsuspecting software developer may incorporate a malicious package into their code, rendering it harmful. Deployment of this code could then result in compromise and escalation into a full-blown cyber-attack.

Figure 1: Flow diagram depicting the initial stages of an AI Package Hallucination Attack.

For providers of Software-as-a-Service (SaaS) products, this attack vector may represent an even greater risk. Such organizations may have a higher proportion of employed software developers than other organizations of comparable size. A threat actor, therefore, could utilize this attack vector as part of a supply chain attack, whereby a malicious payload becomes incorporated into trusted software and is then distributed to multiple customers. This type of attack could have severe consequences including data loss, the downtime of critical systems, and reputational damage.

How could Darktrace detect an AI Package Hallucination Attack?

In June 2023, Darktrace introduced a range of DETECT™ and RESPOND™ models designed to identify the use of generative AI tools within customer environments, and to autonomously perform inhibitive actions in response to such detections. These models will trigger based on connections to endpoints associated with generative AI tools, as such, Darktrace’s detection of an AI Package Hallucination Attack would likely begin with the breaching of one of the following DETECT models:

  • Compliance / Anomalous Upload to Generative AI
  • Compliance / Beaconing to Rare Generative AI and Generative AI
  • Compliance / Generative AI

Should generative AI tool use not be permitted by an organization, the Darktrace RESPOND model ‘Antigena / Network / Compliance / Antigena Generative AI Block’ can be activated to autonomously block connections to endpoints associated with generative AI, thus preventing an AI Package Hallucination attack before it can take hold.

Once a malicious package has been recommended, it may be downloaded from GitHub, a platform and cloud-based service used to store and manage code. Darktrace DETECT is able to identify when a device has performed a download from an open-source repository such as GitHub using the following models:

  • Device / Anomalous GitHub Download
  • Device / Anomalous Script Download Followed By Additional Packages

Whatever goal the malicious package has been designed to fulfil will determine the next stages of the attack. Due to their highly flexible nature, AI package hallucinations could be used as an attack vector to deliver a large variety of different malware types.

As GitHub is a commonly used service by software developers and IT professionals alike, traditional security tools may not alert customer security teams to such GitHub downloads, meaning malicious downloads may go undetected. Darktrace’s anomaly-based approach to threat detection, however, enables it to recognize subtle deviations in a device’s pre-established pattern of life which may be indicative of an emerging attack.

Subsequent anomalous activity representing the possible progression of the kill chain as part of an AI Package Hallucination Attack could then trigger an Enhanced Monitoring model. Enhanced Monitoring models are high-fidelity indicators of potential malicious activity that are investigated by the Darktrace analyst team as part of the Proactive Threat Notification (PTN) service offered by the Darktrace Security Operation Center (SOC).

Conclusion

Employees are often considered the first line of defense in cyber security; this is particularly true in the face of an AI Package Hallucination Attack.

As the use of generative AI becomes more accessible and an increasingly prevalent tool in an attacker’s toolbox, organizations will benefit from implementing company-wide policies to define expectations surrounding the use of such tools. It is simple, yet critical, for example, for employees to fact check responses provided to them by generative AI tools. All packages recommended by generative AI should also be checked by reviewing non-generated data from either external third-party or internal sources. It is also good practice to adopt caution when downloading packages with very few downloads as it could indicate the package is untrustworthy or malicious.

As of September 2023, ChatGPT Plus and Enterprise users were able to use the tool to browse the internet, expanding the data ChatGPT can access beyond the previous training data cut-off of September 2021 [5]. This feature will be expanded to all users soon [6]. ChatGPT providing up-to-date responses could prompt the evolution of this attack vector, allowing attackers to publish malicious packages which could subsequently be recommended by ChatGPT.

It is inevitable that a greater embrace of AI tools in the workplace will be seen in the coming years as the AI technology advances and existing tools become less novel and more familiar. By fighting fire with fire, using AI technology to identify AI usage, Darktrace is uniquely placed to detect and take preventative action against malicious actors capitalizing on the AI boom.

Credit to Charlotte Thompson, Cyber Analyst, Tiana Kelly, Analyst Team Lead, London, Cyber Analyst

References

[1] https://seo.ai/blog/chatgpt-user-statistics-facts

[2] https://darktrace.com/news/darktrace-addresses-generative-ai-concerns

[3] https://darktrace.com/news/darktrace-email-defends-organizations-against-evolving-cyber-threat-landscape

[4] https://vulcan.io/blog/ai-hallucinations-package-risk?nab=1&utm_referrer=https%3A%2F%2Fwww.google.com%2F

[5] https://twitter.com/OpenAI/status/1707077710047216095

[6] https://www.reuters.com/technology/openai-says-chatgpt-can-now-browse-internet-2023-09-27/

Inside the SOC
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Written by
Charlotte Thompson
Cyber Analyst
Written by
Tiana Kelly
Deputy Team Lead, London & Cyber Analyst

More in this series

No items found.

Blog

/

AI

/

July 17, 2025

Introducing the AI Maturity Model for Cybersecurity

AI maturity model for cybersecurityDefault blog imageDefault blog image

AI adoption in cybersecurity: Beyond the hype

Security operations today face a paradox. On one hand, artificial intelligence (AI) promises sweeping transformation from automating routine tasks to augmenting threat detection and response. On the other hand, security leaders are under immense pressure to separate meaningful innovation from vendor hype.

To help CISOs and security teams navigate this landscape, we’ve developed the most in-depth and actionable AI Maturity Model in the industry. Built in collaboration with AI and cybersecurity experts, this framework provides a structured path to understanding, measuring, and advancing AI adoption across the security lifecycle.

Overview of AI maturity levels in cybersecurity

Why a maturity model? And why now?

In our conversations and research with security leaders, a recurring theme has emerged:

There’s no shortage of AI solutions, but there is a shortage of clarity and understanding of AI uses cases.

In fact, Gartner estimates that “by 2027, over 40% of Agentic AI projects will be canceled due to escalating costs, unclear business value, or inadequate risk controls. Teams are experimenting, but many aren’t seeing meaningful outcomes. The need for a standardized way to evaluate progress and make informed investments has never been greater.

That’s why we created the AI Security Maturity Model, a strategic framework that:

  • Defines five clear levels of AI maturity, from manual processes (L0) to full AI Delegation (L4)
  • Delineating the outcomes derived between Agentic GenAI and Specialized AI Agent Systems
  • Applies across core functions such as risk management, threat detection, alert triage, and incident response
  • Links AI maturity to real-world outcomes like reduced risk, improved efficiency, and scalable operations

[related-resource]

How is maturity assessed in this model?

The AI Maturity Model for Cybersecurity is grounded in operational insights from nearly 10,000 global deployments of Darktrace's Self-Learning AI and Cyber AI Analyst. Rather than relying on abstract theory or vendor benchmarks, the model reflects what security teams are actually doing, where AI is being adopted, how it's being used, and what outcomes it’s delivering.

This real-world foundation allows the model to offer a practical, experience-based view of AI maturity. It helps teams assess their current state and identify realistic next steps based on how organizations like theirs are evolving.

Why Darktrace?

AI has been central to Darktrace’s mission since its inception in 2013, not just as a feature, but the foundation. With over a decade of experience building and deploying AI in real-world security environments, we’ve learned where it works, where it doesn’t, and how to get the most value from it. This model reflects that insight, helping security leaders find the right path forward for their people, processes, and tools

Security teams today are asking big, important questions:

  • What should we actually use AI for?
  • How are other teams using it — and what’s working?
  • What are vendors offering, and what’s just hype?
  • Will AI ever replace people in the SOC?

These questions are valid, and they’re not always easy to answer. That’s why we created this model: to help security leaders move past buzzwords and build a clear, realistic plan for applying AI across the SOC.

The structure: From experimentation to autonomy

The model outlines five levels of maturity :

L0 – Manual Operations: Processes are mostly manual with limited automation of some tasks.

L1 – Automation Rules: Manually maintained or externally-sourced automation rules and logic are used wherever possible.

L2 – AI Assistance: AI assists research but is not trusted to make good decisions. This includes GenAI agents requiring manual oversight for errors.

L3 – AI Collaboration: Specialized cybersecurity AI agent systems  with business technology context are trusted with specific tasks and decisions. GenAI has limited uses where errors are acceptable.

L4 – AI Delegation: Specialized AI agent systems with far wider business operations and impact context perform most cybersecurity tasks and decisions independently, with only high-level oversight needed.

Each level reflects a shift, not only in technology, but in people and processes. As AI matures, analysts evolve from executors to strategic overseers.

Strategic benefits for security leaders

The maturity model isn’t just about technology adoption it’s about aligning AI investments with measurable operational outcomes. Here’s what it enables:

SOC fatigue is real, and AI can help

Most teams still struggle with alert volume, investigation delays, and reactive processes. AI adoption is inconsistent and often siloed. When integrated well, AI can make a meaningful difference in making security teams more effective

GenAI is error prone, requiring strong human oversight

While there is a lot of hype around GenAI agentic systems, teams will need to account for inaccuracy and hallucination in Agentic GenAI systems.

AI’s real value lies in progression

The biggest gains don’t come from isolated use cases, but from integrating AI across the lifecycle, from preparation through detection to containment and recovery.

Trust and oversight are key initially but evolves in later levels

Early-stage adoption keeps humans fully in control. By L3 and L4, AI systems act independently within defined bounds, freeing humans for strategic oversight.

People’s roles shift meaningfully

As AI matures, analyst roles consolidate and elevate from labor intensive task execution to high-value decision-making, focusing on critical, high business impact activities, improving processes and AI governance.

Outcome, not hype, defines maturity

AI maturity isn’t about tech presence, it’s about measurable impact on risk reduction, response time, and operational resilience.

[related-resource]

Outcomes across the AI Security Maturity Model

The Security Organization experiences an evolution of cybersecurity outcomes as teams progress from manual operations to AI delegation. Each level represents a step-change in efficiency, accuracy, and strategic value.

L0 – Manual Operations

At this stage, analysts manually handle triage, investigation, patching, and reporting manually using basic, non-automated tools. The result is reactive, labor-intensive operations where most alerts go uninvestigated and risk management remains inconsistent.

L1 – Automation Rules

At this stage, analysts manage rule-based automation tools like SOAR and XDR, which offer some efficiency gains but still require constant tuning. Operations remain constrained by human bandwidth and predefined workflows.

L2 – AI Assistance

At this stage, AI assists with research, summarization, and triage, reducing analyst workload but requiring close oversight due to potential errors. Detection improves, but trust in autonomous decision-making remains limited.

L3 – AI Collaboration

At this stage, AI performs full investigations and recommends actions, while analysts focus on high-risk decisions and refining detection strategies. Purpose-built agentic AI systems with business context are trusted with specific tasks, improving precision and prioritization.

L4 – AI Delegation

At this stage, Specialized AI Agent Systems performs most security tasks independently at machine speed, while human teams provide high-level strategic oversight. This means the highest time and effort commitment activities by the human security team is focused on proactive activities while AI handles routine cybersecurity tasks

Specialized AI Agent Systems operate with deep business context including impact context to drive fast, effective decisions.

Join the webinar

Get a look at the minds shaping this model by joining our upcoming webinar using this link. We’ll walk through real use cases, share lessons learned from the field, and show how security teams are navigating the path to operational AI safely, strategically, and successfully.

Continue reading
About the author
Ashanka Iddya
Senior Director, Product Marketing

Blog

/

Cloud

/

July 17, 2025

Forensics or Fauxrensics: Five Core Capabilities for Cloud Forensics and Incident Response

people working and walking in officeDefault blog imageDefault blog image

The speed and scale at which new cloud resources can be spun up has resulted in uncontrolled deployments, misconfigurations, and security risks. It has had security teams racing to secure their business’ rapid migration from traditional on-premises environments to the cloud.

While many organizations have successfully extended their prevention and detection capabilities to the cloud, they are now experiencing another major gap: forensics and incident response.

Once something bad has been identified, understanding its true scope and impact is nearly impossible at times. The proliferation of cloud resources across a multitude of cloud providers, and the addition of container and serverless capabilities all add to the complexities. It’s clear that organizations need a better way to manage cloud incident response.

Security teams are looking to move past their homegrown solutions and open-source tools to incorporate real cloud forensics capabilities. However, with the increased buzz around cloud forensics, it can be challenging to decipher what is real cloud forensics, and what is “fauxrensics.”

This blog covers the five core capabilities that security teams should consider when evaluating a cloud forensics and incident response solution.

[related-resource]

1. Depth of data

There have been many conversations among the security community about whether cloud forensics is just log analysis. The reality, however, is that cloud forensics necessitates access to a robust dataset that extends far beyond traditional log data sources.

While logs provide valuable insights, a forensics investigation demands a deeper understanding derived from multiple data sources, including disk, network, and memory, within the cloud infrastructure. Full disk analysis complements log analysis, offering crucial context for identifying the root cause and scope of an incident.

For instance, when investigating an incident involving a Kubernetes cluster running on an EC2 instance, access to bash history can provide insights into the commands executed by attackers on the affected instance, which would not be available through cloud logs alone.

Having all of the evidence in one place is also a capability that can significantly streamline investigations, unifying your evidence be it disk images, memory captures or cloud logs, into a single timeline allowing security teams to reconstruct an attacks origin, path and impact far more easily. Multi–cloud environments also require platforms that can support aggregating data from many providers and services into one place. Doing this enables more holistic investigations and reduces security blind spots.

There is also the importance of collecting data from ephemeral resources in modern cloud and containerized environments. Critical evidence can be lost in seconds as resources are constantly spinning up and down, so having the ability to capture this data before its gone can be a huge advantage to security teams, rather than having to figure out what happened after the affected service is long gone.

darktrace / cloud, cado, cloud logs, ost, and memory information. value of cloud combined analysis

2. Chain of custody

Chain of custody is extremely critical in the context of legal proceedings and is an essential component of forensics and incident response. However, chain of custody in the cloud can be extremely complex with the number of people who have access and the rise of multi-cloud environments.

In the cloud, maintaining a reliable chain of custody becomes even more complex than it already is, due to having to account for multiple access points, service providers and third parties. Having automated evidence tracking is a must. It means that all actions are logged, from collection to storage to access. Automation also minimizes the chance of human error, reducing the risk of mistakes or gaps in evidence handling, especially in high pressure fast moving investigations.

The ability to preserve unaltered copies of forensic evidence in a secure manner is required to ensure integrity throughout an investigation. It is not just a technical concern, its a legal one, ensuring that your evidence handling is documented and time stamped allows it to stand up to court or regulatory review.

Real cloud forensics platforms should autonomously handle chain of custody in the background, recording and safeguarding evidence without human intervention.

3. Automated collection and isolation

When malicious activity is detected, the speed at which security teams can determine root cause and scope is essential to reducing Mean Time to Response (MTTR).

Automated forensic data collection and system isolation ensures that evidence is collected and compromised resources are isolated at the first sign of malicious activity. This can often be before an attacker has had the change to move latterly or cover their tracks. This enables security teams to prevent potential damage and spread while a deeper-dive forensics investigation takes place. This method also ensures critical incident evidence residing in ephemeral environments is preserved in the event it is needed for an investigation. This evidence may only exist for minutes, leaving no time for a human analyst to capture it.

Cloud forensics and incident response platforms should offer the ability to natively integrate with incident detection and alerting systems and/or built-in product automation rules to trigger evidence capture and resource isolation.

4. Ease of use

Security teams shouldn’t require deep cloud or incident response knowledge to perform forensic investigations of cloud resources. They already have enough on their plates.

While traditional forensics tools and approaches have made investigation and response extremely tedious and complex, modern forensics platforms prioritize usability at their core, and leverage automation to drastically simplify the end-to-end incident response process, even when an incident spans multiple Cloud Service Providers (CSPs).

Useability is a core requirement for any modern forensics platform. Security teams should not need to have indepth knowledge of every system and resource in a given estate. Workflows, automation and guidance should make it possible for an analyst to investigate whatever resource they need to.

Unifying the workflow across multiple clouds can also save security teams a huge amount of time and resources. Investigations can often span multiple CSP’s. A good security platform should provide a single place to search, correlate and analyze evidence across all environments.

Offering features such as cross cloud support, data enrichment, a single timeline view, saved search, and faceted search can help advanced analysts achieve greater efficiency, and novice analysts are able to participate in more complex investigations.

5. Incident preparedness

Incident response shouldn't just be reactive. Modern security teams need to regularly test their ability to acquire new evidence, triage assets and respond to threats across both new and existing resources, ensuring readiness even in the rapidly changing environments of the cloud.  Having the ability to continuously assess your incident response and forensics workflows enables you to rapidly improve your processes and identify and mitigate any gaps identified that could prevent the organization from being able to effectively respond to potential threats.

Real forensics platforms deliver features that enable security teams to prepare extensively and understand their shortcomings before they are in the heat of an incident. For example, cloud forensics platforms can provide the ability to:

  • Run readiness checks and see readiness trends over time
  • Identify and mitigate issues that could prevent rapid investigation and response
  • Ensure the correct logging, management agents, and other cloud-native tools are appropriately configured and operational
  • Ensure that data gathered during an investigation can be decrypted
  • Verify that permissions are aligned with best practices and are capable of supporting incident response efforts

Cloud forensics with Darktrace

Darktrace delivers a proactive approach to cyber resilience in a single cybersecurity platform, including cloud coverage. Darktrace / CLOUD is a real time Cloud Detection and Response (CDR) solution built with advanced AI to make cloud security accessible to all security teams and SOCs. By using multiple machine learning techniques, Darktrace brings unprecedented visibility, threat detection, investigation, and incident response to hybrid and multi-cloud environments.

Darktrace’s cloud offerings have been bolstered with the acquisition of Cado Security Ltd., which enables security teams to gain immediate access to forensic-level data in multi-cloud, container, serverless, SaaS, and on-premises environments.

[related-resource]

Continue reading
About the author
Calum Hall
Technical Content Researcher
Your data. Our AI.
Elevate your network security with Darktrace AI