Blog
/

Inside the SOC

/
October 30, 2023

Exploring AI Threats: Package Hallucination Attacks

Default blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog image
30
Oct 2023
Learn how malicious actors exploit errors in generative AI tools to launch packet attacks. Read how Darktrace products detect and prevent these threats!

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.
Author
Charlotte Thompson
Cyber Analyst
Tiana Kelly
Deputy Team Lead, London & Cyber Analyst
Book a 1-1 meeting with one of our experts
Share this article

More in this series

No items found.

Blog

/

November 28, 2024

/

Cloud

Cloud security: addressing common CISO challenges with advanced solutions

Default blog imageDefault blog image

Cloud adoption is a cornerstone of modern business with its unmatched potential for scalability, cost efficiency, flexibility, and net-zero targets around sustainability. However, as organizations migrate more workloads, applications, and sensitive data to the cloud it introduces more complex challenges for CISO’s. Let’s dive into the most pressing issues keeping them up at night—and how Darktrace / CLOUD provides a solution for each.

1. Misconfigurations: The Silent Saboteur

Misconfigurations remain the leading cause of cloud-based data breaches. In 2023 alone over 80%  of data breaches involved data stored in the cloud.1  Think open storage buckets or overly permissive permissions; seemingly minor errors that are easily missed and can snowball into major disasters. The fallout of breaches can be costly—both financially and reputationally.

How Darktrace / CLOUD Helps:

Darktrace / CLOUD continuously monitors your cloud asset configurations, learning your environment and using these insights to flag potential misconfigurations. New scans are triggered when changes take place, then grouped and prioritised intelligently, giving you an evolving and prioritised view of vulnerabilities, best practice and mitigation strategies.

2. Hybrid Environments: The Migration Maze

Many organizations are migrating to the cloud, but hybrid setups (where workloads span both on-premises and cloud environments) create unique challenges and visibility gaps which significantly increase complexity. More traditional and most cloud native security tooling struggles to provide adequate monitoring for these setups.

How Darktrace / CLOUD Helps:

Provides the ability to monitor runtime activity for both on-premises and cloud workloads within the same user interface. By leveraging the right AI solution across this diverse data set, we understand the behaviour of your on-premises workloads and how they interact with cloud systems, spotting unusual connectivity or data flow activity during and after the migration process.

This unified visibility enables proactive detection of anomalies, ensures seamless monitoring across hybrid environments, and provides actionable insights to mitigate risks during and after the migration process.

3. Securing Productivity Suites: The Last Mile

Cloud productivity suites like Microsoft 365 (M365) are essential for modern businesses and are often the first step for an organization on a journey to Infrastructure as a Service (IaaS) or Platform as a Service (PaaS) use cases. They also represent a prime target for attackers. Consider a scenario where an attacker gains access to an M365 account, and proceeds to; access sensitive emails, downloading files from SharePoint, and impersonating the user to send phishing emails to internal employees and external partners. Without a system to detect these behaviours, the attack may go unnoticed until significant damage is done.

How Darktrace helps:

Darktrace’s Active AI platform integrates with M365 and establishes an understanding of normal business activity, enabling the detection of abnormalities across its suite including Email, SharePoint and Teams. By identifying subtle deviations in behaviour, such as:

   •    Unusual file accesses

   •    Anomalous login attempts from unexpected locations or devices.

   •    Suspicious email forwarding rules created by compromised accounts.

Darktrace’s Autonomous Response can act precisely to block malicious actions, by disabling compromised accounts and containing threats before they escalate. Precise actions also ensure that critical business operations are maintained even when a response is triggered.  

4. Agent Fatigue: The Visibility Struggle

To secure cloud environments, visibility is critical. If you don’t know what’s there, how can you secure it? Many solutions require agents to be deployed on every server, workload, and endpoint. But managing and deploying agents across sprawling hybrid environments can be both complex and time-consuming when following change controls, and especially as cloud resources scale dynamically.

How Darktrace / CLOUD Helps:

Darktrace reduces or eliminates the need for widespread agent deployment. Its agentless by default, integrating directly with cloud environments and providing instant visibility without the operational headache. Darktrace ensures coverage with minimal friction. By intelligently graphing the relationships between assets and logically grouping your deployed Cloud resources, you are equipped with real-time visibility to quickly understand and protect your environment.

So why Darktrace / CLOUD?

Darktrace’s Self-Learning AI redefines cloud security by adapting to your unique environment, detecting threats as they emerge, and responding in real-time. From spotting misconfigurations to protecting productivity suites and securing hybrid environments. Darktrace / CLOUD simplifies cloud security challenges without adding operational burdens.

From Chaos to Clarity

Cloud security doesn’t have to be a game of endless whack-a-mole. With Darktrace / CLOUD, CISOs can achieve the visibility, control, and proactive protection they need to navigate today’s complex cloud ecosystems confidently.

[1] https://hbr.org/2024/02/why-data-breaches-spiked-in-2023

Continue reading
About the author
Adam Stevens
Director of Product, Cloud Security

Blog

/

November 28, 2024

/

Thought Leadership

Preparing for 2025: Darktrace's top 10 AI and cybersecurity predictions

Default blog imageDefault blog image

Each year, Darktrace's AI and cybersecurity experts reflect on the events of the past 12 months and predict the trends we expect to shape the cybersecurity landscape in the year ahead. In 2024, we predicted that the global elections, fast-moving AI innovations, and increasingly cloud-based IT environments would be key factors shaping the cyber threat landscape.

Looking ahead to 2025, we expect the total addressable market of cybercrime to expand as attackers add more tactics to their toolkits. Threat actors will continue to take advantage of the volatile geopolitical environment and cybersecurity challenges will increasingly move to new frontiers like space. When it comes to AI, we anticipate the innovation in AI agents in 2024 to pave the way for the rise of multi-agent systems in 2025, creating new challenges and opportunities for cybersecurity professionals and attackers alike.

Here are ten trends to watch for in 2025:

The overall Total Addressable Market (TAM) of cybercrime gets bigger

Cybercrime is a global business, and an increasingly lucrative one, scaling through the adoption of AI and cybercrime-as-a-service. Annual revenue from cybercrime is already estimated to be over $8 trillion, which we’ve found is almost 5x greater than the revenue of the Magnificent Seven stocks. There are a few key factors driving this growth.

The ongoing growth of devices and systems means that existing malware families will continue to be successful. As of October 2024, it’s estimated that more than 5.52 billion people (~67%) have access to the internet and sources estimate 18.8 billion connected devices will be online by the end of 2024. The increasing adoption of AI is poised to drive even more interconnected systems as well as new data centers and infrastructure globally.

At the same time, more sophisticated capabilities are available for low-level attackers – we’ve already seen the trickle-down economic benefits of living off the land, edge infrastructure exploitation, and identity-focused exploitation. The availability of Ransomware-as-a-Service (RaaS) and Malware-as-a-Service (MaaS) make more advanced tactics the norm. The subscription income that these groups can generate enables more adversarial innovation, so attacks are getting faster and more effective with even bigger financial ramifications.

While there has also been an increasing trend in the last year of improved cross-border law enforcement, the efficacy of these efforts remains to be seen as cybercriminal gangs are also getting more resilient and professionalized. They are building better back-up systems and infrastructure as well as more multi-national networks and supply chains.

Security teams need to prepare for the rise of AI agents and multi-agent systems

Throughout 2024, we’ve seen major announcements about advancements in AI agents from the likes of OpenAI, Microsoft, Salesforce, and more. In 2025, we’ll see increasing innovation in and adoption of AI agents as well as the emergence of multi-agent systems (or “agent swarms”), where groups of autonomous agents work together to tackle complex tasks.

The rise of AI agents and multi-agent systems will introduce new challenges in cybersecurity, including new attack vectors and vulnerabilities. Security teams need to think about how to protect these systems to prevent data poisoning, prompt injection, or social engineering attacks.

One benefit of multi-agent systems is that agents can autonomously communicate, collaborate, and interact. However without clear and distinct boundaries and explicit permissions, this can also pose a major data privacy risk and avenue for manipulation. These issues cannot be addressed by traditional application testing alone. We must ensure these systems are secure by design, where robust protective mechanisms and data guardrails are built into the foundations.

Threat actors will be the earliest adopters of AI agents and multi-agent systems

We’ve already seen how quickly threat actors have been able to adopt generative AI for tasks like email phishing and reconnaissance. The next frontier for threat actors will be AI agents and multi-agent systems that are specialized in autonomous tasks like surveillance, initial access brokering, privilege escalation, vulnerability exploitation, data summarization for smart exfiltration, and more. Because they have no concern for safe, secure, accurate, and responsible use, adversaries will adopt these systems faster than cyber defenders.

We could also start to see use cases emerge for multi-agent systems in cyber defense – with potential for early use cases in incident response, application testing, and vulnerability discovery. On the whole, security teams will be slower to adopt these systems than adversaries because of the need to put in place proper security guardrails and build trust over time.

There is heightened supply chain risk for Large Language Models (LLMs)

Training LLMs requires a lot of data, and many experts have warned that world is running out of quality data for that training. As a result, there will be an increasing reliance on synthetic data, which can introduce new issues of accuracy and efficacy. Moreover, data supply chain risks will be an Achilles heel for organizations, with the potential interjection of vulnerabilities through the data and machine learning providers that they rely on. Poisoning one data set could have huge trickle-down impacts across many different systems. Data security will be paramount in 2025.

The race to identify software vulnerabilities intensifies

The time it takes for threat actors to exploit newly published CVEs is getting shorter, giving defenders an even smaller window to apply patches and remediations. A 2024 report from Cloudflare found that threat actors quickly weaponized proof of concept exploits in attacks as quickly as 22 minutes after the exploits were made public.

At the same time, 2024 also saw the first reports from researchers across academia and the tech industry using AI for vulnerability discovery in real-world code. With threat actors getting faster at exploiting vulnerabilities, defenders will need to use AI to identify vulnerabilities in their software stack and to help identify and prioritize remediations and patches.

Insider threat risks will force organizations to evolve zero trust strategies

In 2025, an increasingly volatile geopolitical situation and the intensity of the AI race will make insider threats an even bigger risk for businesses, forcing organizations to expand zero-trust strategies. The traditional zero-trust model provides protection from external threats to an organization’s network by requiring continuous verification of the devices and users attempting to access critical business systems, services, and information from multiple sources. However, as we have seen in the more recent Jack Teixeira case, malicious insiders can still do significant damage to an organization within their approved and authenticated boundary.

To circumvent the remaining security gaps in a zero-trust architecture and mitigate increasing risk of insider threats, organizations will need to integrate a behavioral understanding dimension to their zero-trust approaches. The zero-trust best practice of “never trust, always verify” needs to evolve to become “never trust, always verify, and continuously monitor.”

Identity remains an expensive problem for businesses

2024 saw some of the biggest and costliest attacks – all because the attacker had access to compromised credentials. Essentially, they had the key to the front door. Businesses still struggle with identity and access management (IAM), and it’s getting more complex now that we’re in the middle of a massive Software-as-a-Service (SaaS) migration driven by increasing rates of AI and cloud use across businesses.

This challenge is going to be exacerbated in 2025 by a few global and business factors. First, there is an increasing push for digital identities, such as the rollout of the EU Digital Identity Framework that is underway, which could introduce additional attack vectors. As they scale, businesses are turning more and more to centralized identity and access solutions with decentralized infrastructure and relying on SaaS and application-native security.

Increasing vulnerabilities at the edge

During the COVID-19 pandemic, many organizations had to stand-up remote access solutions quickly – in a matter of days or weeks – without the high level of due diligence that they require to be fully secured. In 2025, we expect to see continued fall-out as these quickly spun-up solutions start to present genuine vulnerability to businesses. We’ve already seen this start to play out in 2024 with the mass-exploitation of internet-edge devices like firewalls and VPN gateway products.

By July 2024, Darktrace’s threat research team observed that the most widely exploited edge infrastructure devices were those related to Ivanti Connect Secure, JetBrains TeamCity, FortiClient Enterprise Management Server, and Palo Alto Networks PAN-OS. Across the industry, we’ve already seen many zero days and vulnerabilities exploiting these internet-connected devices, which provide inroads into the network and store/cache credentials and passwords of other users that are highly valuable for threat actors.

Hacking Operational Technology (OT) gets easier

Hacking OT is notoriously complex – causing damage requires an intimate knowledge of the specific systems being targeted and historically was the reserve of nation states. But as OT has become more reliant and integrated with IT systems, attackers have stumbled on ways to cause disruption without having to rely on the sophisticated attack-craft normally associated with nation-state groups. That’s why some of the most disruptive attacks of the last year have come from hacktivist and financially-motivated criminal gangs – such as the hijacking of internet-exposed Programmable Logic Controllers (PLCs) by anti-Israel hacking groups and ransomware attacks resulting in the cancellation of hospital operations.  

In 2025, we expect to see an increase in cyber-physical disruption caused by threat groups motivated by political ideology or financial gain, bringing the OT threat landscape closer in complexity and scale to that of the IT landscape. The sectors most at risk are those with a strong reliance on IoT sensors, including healthcare, transportation, and manufacturing sectors.

Securing space infrastructure and systems becomes a critical imperative

The global space industry is growing at an incredibly fast pace, and 2025 is on track to be another record-breaking year for spaceflight with major missions and test flights planned by NASA, ESA, CNSA as well as the expected launch of the first commercial space station from Vast and programs from Blue Origin, Amazon and more. Research from Analysis Mason suggests that 38,000 additional satellites will be built and launched by 2033 and the global space industry revenue will reach $1.7 trillion by 2032. Space has also been identified as a focus area for the incoming US administration.

In 2025, we expect to see new levels of tension emerge as private and public infrastructure increasingly intersect in space, shining a light on the lack of agreed upon cyber norms and the increasing challenge of protecting complex and remote space systems against modern cyber threats.  Historically focused on securing earth-bound networks and environments, the space industry will face challenges as post-orbit threats rise, with satellites moving up the target list.

The EU’s NIS2 Directive now recognizes the space sector as an essential entity that is subject to its most strict cybersecurity requirements. Will other jurisdictions follow suit? We expect global debates about cyber vulnerabilities in space to come to the forefront as we become more reliant on space-based technology.

Preparing for the future

Whatever 2025 brings, Darktrace is committed to providing robust cybersecurity leadership and solutions to enterprises around the world. Our team of subject matter experts will continue to monitor emerging threat trends, advising both our customers and our product development teams.

And for day-to-day security, our multi-layered AI cybersecurity platform can protect against all types of threats, whether they are known, unknown, entirely novel, or powered by AI. It accomplishes this by learning what is normal for your unique organization, therefore identifying unusual and suspicious behavior at machine speed, regardless of existing rules and signatures. In this way, organizations with Darktrace can be ready for any developments in the cybersecurity threat landscape that the new year may bring.

Discover more about Darktrace's predictions on the AI and cybersecurity landscape for 2025 by joining the upcoming webinar on December 12, 2024 at 10:00am EST/3:00pm GMT. Register here.

Continue reading
About the author
The Darktrace Community
Your data. Our AI.
Elevate your network security with Darktrace AI