Technical brief · AI Security

AI security risk landscape

A concise map of AI-specific threats, from prompt injection and data exposure to model risks and agent security.

On this page
  1. AI-specific threats
  2. LLM security
  3. Prompt injection
  4. Data exposure
  5. Model-related risks
  6. AI agent security
  7. Governance considerations

Educational resource. This is general guidance. It is not an official certification, a regulatory interpretation or legal advice.

AI-specific threats

AI systems inherit all the usual application security risks and add some of their own. Shared references include the OWASP Top 10 for LLM Applications, MITRE ATLAS and NIST’s adversarial machine learning taxonomy. This field changes quickly, so treat any single list as a starting point.

LLM security

The 2025 OWASP list covers prompt injection, sensitive information disclosure, supply chain, data and model poisoning, improper output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, misinformation and unbounded consumption.

Prompt injection

Models receive instructions and data as the same stream of text, so content can be crafted to override intended behaviour. Indirect injection hides instructions in documents or web pages that the system later processes. Filters help but are not sufficient. Design on the assumption that some attempts will succeed, and limit what a compromised response can do.

Data exposure

Sensitive data can leak through responses, logs, retrieval indexes that ignore user permissions, and third-party services. Apply access control at the retrieval layer, minimise the data sent to models, and check your vendors’ data handling terms.

Training data can be poisoned, and models or their dependencies can carry supply chain risk. Track where models and datasets come from, and verify them before use. Re-evaluate behaviour after every model update, because results can change without any change to your code.

AI agent security

Agents turn model output into action, which raises the stakes of every risk above. Give each agent its own identity, grant least privilege, require human approval for high-impact actions, cap autonomy and keep logs that attribute each action. OWASP’s agentic applications list covers this area in more depth.

Governance considerations

  • Include AI systems in your existing security review and incident processes.
  • Threat model higher-risk systems before deployment, and again when they change.
  • Assign clear ownership for AI security decisions between security and AI teams.
  • Assess vendors on model provenance, data handling and how they notify you of changes.

Sources

AI securityLLM securityPrompt injectionAI supply chainAI agent security

Share: LinkedInXEmail

Talk to Cognexaa

Knowledge creates confidence. Technology makes it possible.

Talk to our team about how Cognexaa’s platforms can help your organisation approach AI governance, cybersecurity and digital trust with greater structure.

Or email us directly at sales@cognexaa.com