AI risk atlas
Last updated: May 29, 2025
Explore this atlas to understand some of the risks of working with agentic AI, generative AI, and machine learning models.
On this page
Back to top
New and amplified risks of agentic AI
Risks are categorized with one of these tags:
Amplified by agentic AI
Risks that are more severe or likely due to agentic AI.Specific to agentic AI
Risks that are specifically associated with agentic AI.An AI agent is a software entity that employs AI techniques and has agency to act in its environment based on set goals, which means it can decide which actions to perform and has the ability to execute them. Agentic AI systems are software systems that leverage AI agents (together with other components like tools, planners, memory, and datasets), pursue goals, and can operate autonomously.
AI agents can perform three types of actions:
- Take actions that impact the world (physical or digital).
- Consult resources and use tools.
- Decide which process to choose in the selection of resources/tools/other AI agents and select them.
The risks in this section are specific to or amplified by agentic AI. Since recent agents are built on large language models, the generative AI risks in the following section may also be applicable to agentic AI.
Privacy
Sharing IP/PI/confidential information with user
Amplified by agentic AI
Sharing IP/PI/confidential information with tools
Specific to agentic AI
Value alignment
Over- or under-reliance on AI agents
Amplified by agentic AI
Misaligned actions
Amplified by agentic AI
Robustness
Attack on AI agents’ external resources
Specific to agentic AI
Unauthorized use
Amplified by agentic AI
Exploit trust mismatch
Amplified by agentic AI
Function calling hallucination
Specific to agentic AI
Computational inefficiency
Redundant actions
Specific to agentic AI
Governance
Incomplete AI agent evaluation
Amplified by agentic AI
Mitigation and maintenance
Amplified by agentic AI
Lack of AI agent transparency
Amplified by agentic AI
Reproducibility
Specific to agentic AI
Accountability of AI agent actions
Amplified by agentic AI
AI agent compliance
Amplified by agentic AI
Societal impact
Impact on human dignity
Amplified by agentic AI
AI agents' impact on human agency
Amplified by agentic AI
AI agents' impact on jobs
Amplified by agentic AI
AI agents' impact on environment
Amplified by agentic AI
Explainability
Unexplainable and untraceable actions
Amplified by agentic AI
Back to top
All risks
Risks are categorized with one of these tags:
Traditional risk of AI
Established risks of AI that apply to both traditional and generative models.Amplified by generative AI
Risks that are more severe or likely due to generative AI. These risks are also applicable to traditional AI models.Specific to generative AI
Risks that are specifically associated with generative AI models.The risks below describe risks that are applicable to generative AI models and traditional (non-generative) AI models. These risks may also apply to agentic AI, especially in cases where the agent's behavior or output is determined using a generative or traditional AI model.
Training data risks
Fairness
Data bias
Amplified by generative AI
Value alignment
Improper data curation
Amplified by generative AI
Improper retraining
Amplified by generative AI
Robustness
Data poisoning
Traditional risk of AI
Privacy
Personal information in data
Traditional risk of AI
Reidentification
Traditional risk of AI
Data privacy rights alignment
Amplified by generative AI
Transparency
Lack of training data transparency
Amplified by generative AI
Uncertain data provenance
Amplified by generative AI
Data laws
Data acquisition restrictions
Amplified by generative AI
Data usage restrictions
Traditional risk of AI
Data transfer restrictions
Traditional risk of AI
Intellectual property
Confidential information in data
Amplified by generative AI
Data usage rights restrictions
Amplified by generative AI
Inference risks
Accuracy
Poor model accuracy
Amplified by generative AI
Robustness: Model behavior manipulation
Evasion attack
Amplified by generative AI
Extraction attack
Amplified by generative AI
Jailbreaking
Specific to generative AI
Intellectual property
IP information in prompt
Specific to generative AI
Confidential data in prompt
Specific to generative AI
Robustness: Prompt attacks
Prompt injection attack
Specific to generative AI
Prompt leaking
Specific to generative AI
Prompt priming
Specific to generative AI
Context overload attack
Specific to generative AI
Direct instructions attack
Specific to generative AI
Encoded interactions attack
Specific to generative AI
Indirect instructions attack
Specific to generative AI
Social hacking attack
Specific to generative AI
Specialized tokens attack
Specific to generative AI
Privacy
Personal information in prompt
Specific to generative AI
Attribute inference attack
Amplified by generative AI
Membership inference attack
Amplified by generative AI
Output risks
Value alignment
Harmful output
Specific to generative AI
Harmful code generation
Specific to generative AI
Toxic output
Specific to generative AI
Incomplete advice
Specific to generative AI
Over- or under-reliance
Amplified by generative AI
Misuse
Dangerous use
Specific to generative AI
Spreading disinformation
Specific to generative AI
Nonconsensual use
Specific to generative AI
Spreading toxicity
Specific to generative AI
Improper usage
Amplified by generative AI
Non-disclosure
Specific to generative AI
Robustness
Hallucination
Specific to generative AI
Privacy
Exposing personal information
Amplified by generative AI
Intellectual property
Copyright infringement
Specific to generative AI
Revealing confidential information
Amplified by generative AI
Explainability
Unexplainable output
Amplified by generative AI
Unreliable source attribution
Specific to generative AI
Untraceable attribution
Amplified by generative AI
Inaccessible training data
Amplified by generative AI
Non-technical risks
Governance
Lack of data transparency
Amplified by generative AI
Lack of model transparency
Traditional risk of AI
Lack of system transparency
Traditional risk of AI
Incomplete usage definition
Specific to generative AI
Incorrect risk testing
Amplified by generative AI
Unrepresentative risk testing
Amplified by generative AI
Lack of testing diversity
Amplified by generative AI
Legal compliance
Model usage rights restrictions
Traditional risk of AI
Legal accountability
Amplified by generative AI
Generated content ownership and IP
Specific to generative AI
Societal impact
Impact on the environment
Amplified by generative AI
Impact on affected communities
Traditional risk of AI
Human exploitation
Amplified by generative AI
Impact on Jobs
Amplified by generative AI
AI agents' Impact on human agency
Amplified by generative AI
Impact on cultural diversity
Specific to generative AI
Impact on education: bypassing learning
Specific to generative AI
Impact on education: plagiarism
Specific to generative AI
Was the topic helpful?
0/1000