MIT AI Risk Repository · domain 2: Privacy & Security

2.2 AI system security vulnerabilities and attacks

Vulnerabilities in AI systems, software development toolchains, and hardware that can be exploited, resulting in unauthorized access, data and privacy breaches, or system manipulation causing unsafe outputs or behavior.

Risk entries
112
Frameworks citing it
12
Recorded incidents
24
Incidents since 2020
20
Causal entity (risk entries)
Causal entity (risk entries) 87 0 Human: 87 Human 87 Other: 18 Other 18 AI: 6 AI 6 Not coded: 1 Not coded 1
Causal entity (risk entries)
LabelValue
Human87
Other18
AI6
Not coded1
Intent (risk entries)
Intent (risk entries) 83 0 Intentional: 83 Intentional 83 Unintentional: 16 Unintentional 16 Other: 12 Other 12 Not coded: 1 Not coded 1
Intent (risk entries)
LabelValue
Intentional83
Unintentional16
Other12
Not coded1
Timing (risk entries)
Timing (risk entries) 64 0 Post-deployment: 64 Post-deployment 64 Pre-deployment: 25 Pre-deployment 25 Other: 22 Other 22 Not coded: 1 Not coded 1
Timing (risk entries)
LabelValue
Post-deployment64
Pre-deployment25
Other22
Not coded1
Recorded incidents per yearIncident date; current year partial
Recorded incidents per year 7 0 2017: 3 2017 3 2019: 1 2019 1 2022: 2 2022 2 2023: 3 2023 3 2024: 3 2024 3 2025: 5 2025 5 2026: 7 2026 7
Recorded incidents per year
LabelValue
20173
20191
20222
20233
20243
20255
20267
Entries by levelRisk categories, subcategories and additional evidence coded to this subdomain
Entries by level 96 0 Risk Category: 16 Risk Category 16 Risk Sub-Category: 96 Risk Sub-Category 96
Entries by level
LabelValue
Risk Category16
Risk Sub-Category96
  • Evasion Attacks

    "Evasion attacks [145] target to cause significant shifts in model’s prediction via adding perturbations in the test samples to build adversarial examples. In specific, the perturbations can be implem...

    Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024) · Human · Intentional · Pre-deployment

  • Adversarial Prompts

    "Engineering an adversarial input to elicit an undesired model behavior, which pose a clear attack intention"

    Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024) · Human · Intentional · Post-deployment

  • Goal Hijacking

    "Goal hijacking is a type of primary attack in prompt injection [58]. By injecting a phrase like “Ignore the above instruction and do ...” in the input, the attack could hijack the original goal of th...

    Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024) · Human · Intentional · Post-deployment

  • One-step Jailbreaks

    "One-step jailbreaks. One-step jailbreaks commonly involve direct modifications to the prompt itself, such as setting role-playing scenarios or adding specific descriptions to prompts [14], [52], [67]...

    Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024) · Human · Intentional · Post-deployment

  • Multi-step Jailbreaks

    "Multi-step jailbreaks. Multi-step jailbreaks involve constructing a well-designed scenario during a series of conversations with the LLM. Unlike one-step jailbreaks, multi-step jailbreaks usually gui...

    Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024) · Human · Intentional · Post-deployment

  • Prompt Leaking

    "Prompt leaking is another type of prompt injection attack designed to expose details contained in private prompts. According to [58], prompt leaking is the act of misleading the model to print the pr...

    Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024) · Human · Intentional · Post-deployment

  • Security

    "How to design AGIs that are robust to adversaries and adversarial environ- ments? This involves building sandboxed AGI protected from adversaries (Berkeley), and agents that are robust to adversarial...

    AGI Safety Literature Review (Everitt2018 ) · Human · Unintentional · Pre-deployment

  • Adversarial AI (General)

    "Adversarial AI refers to a class of attacks that exploit vulnerabilities in machine-learning (ML) models. This class of misuse exploits vulnerabilities introduced by the AI assistant itself and is a...

    The Ethics of Advanced AI Assistants (Gabriel2024) · Other · Intentional · Post-deployment

  • Adversarial AI: Circumvention of Technical Security Measures

    "The technical measures to mitigate misuse risks of advanced AI assistants themselves represent a new target for attack. An emerging form of misuse of general-purpose advanced AI assistants exploits v...

    The Ethics of Advanced AI Assistants (Gabriel2024) · Other · Intentional · Post-deployment

  • Adversarial AI: Prompt Injections

    "Prompt injections represent another class of attacks that involve the malicious insertion of prompts or requests in LLM-based interactive systems, leading to unintended actions or disclosure of sensi...

    The Ethics of Advanced AI Assistants (Gabriel2024) · Other · Intentional · Post-deployment

  • Security harms (cybersecurity)

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Not coded · Not coded · Not coded

  • Data-related (Difficulty filtering large web scrapes or large scale web datasets)

    "A large scale “scraping” of web data for training datasets increases vulnerability to data poisoning, backdoor attacks, and the inclusion of inaccurate or toxic data [76, 28, 48]. With a large datase...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Unintentional · Pre-deployment

  • Data-related (Insufficient quality control in data collection process)

    "A lack of standardized methods and sufficient infrastructure, including the absence of quality control processes for collecting data, especially for high-stakes domains and benchmarks, can affect the...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Unintentional · Pre-deployment

  • Training-related (Adversarial examples)

    "Adversarial examples [198, 83] refer to data that are designed to fool an AI model by inducing unintended behavior. They do this by exploiting spurious correlations learned by the model. They are par...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Intentional · Post-deployment

  • Training-related (Robustness certificates can be exploited to attack the models)

    "The knowledge of robustness certificates, including the area of the region for which model predictions are certified to be robust, can be used by an adversary to efficiently craft attacks that succee...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Intentional · Other

  • Fine-tuning related (Fine-tuning dataset poisoning)

    "A deployer can poison the dataset used during the fine-tuning process [98] to induce specific, often malicious, behaviors in a model. This can be performed without having access to the model’s weight...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Intentional · Pre-deployment

  • Fine-tuning related (Poisoning models during instruction tuning)

    "AI models can be poisoned during instruction tuning when models are tuned using pairs of instructions and desired outputs. Poisoning in instruction tuning can be achieved with a lower number of compr...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Intentional · Pre-deployment

  • Misuse of interpretability techniques

    "Interpretability techniques, by enabling a better understanding of the model, could potentially be used for harmful purposes. For example, mechanistic inter- pretability could be used to identify neu...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Intentional · Other

  • Adversarial attacks targeting explainable AI techniques

    "Adversarial attacks can affect not only the model’s output but also its corresponding explanation. Current adversarial optimization techniques can intro- duce imperceptible noise to the input image,...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Intentional · Other

  • Jailbreak of a model to subvert intended behavior

    "A jailbreak is a type of adversarial input to the model (during deployment) re- sulting in model behavior deviating from intended use. Jailbreaks may be gen- erated automatically in a “white box” set...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Intentional · Post-deployment

  • Jailbreak of a multimodal model

    "Current generation multimodal (e.g., vision and language) GPAI models are vulnerable to adversarial jailbreak attacks. These attacks can be used to automatically induce a model to produce an arbitrar...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Intentional · Post-deployment

  • Transferable adversarial attacks from open to closed-source mod- els

    "In some cases, an adversarial attack developed for an open-weights and open- source model (where the weights and architecture are known - a “white box” attack) can be transferable to closed-source mo...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Intentional · Post-deployment

  • Backdoors or trojan attacks in GPAI models

    "Backdoors can be inserted into GPAI models during their training or fine-tuning, to be exploited during deployment [185, 118]. Attackers inserting the backdoor can be the GPAI model provider themselv...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Intentional · Pre-deployment

  • Text encoding-based attacks

    "Various new or existing text encodings, such as Base64, can be employed to craft jailbreak attacks that bypass safety training [13]. Low-resource language inputs also appear more likely to circumvent...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Human · Intentional · Post-deployment

  • Vulnerabilities arising from additional modalities in multimodal models

    "Additional modalities can introduce new attack vectors in multimodal models as well as expand the scope of the previous attacks, ranging from jailbreaking to poisoning [13]. Typically, different moda...

    Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024) · Other · Other · Other