Natural Habitat.
AI Security: Data, Permissions and Risks
Using AI securely requires control over the information sent to the model, the use of its responses and access to tools capable of taking action.
The most important risks depend on the particular architecture and system permissions, not simply on the presence of generative functionality.
Confidential data and retention
OWASP identifies sensitive information disclosure as category LLM02:2025. Personal information, financial records, business documents and credentials may be exposed. Limiting input data and reviewing storage rules are fundamental safeguards.
During an AI security audit, establish which systems can see the data, how long it is stored and whether it can be deleted. Hiding part of the interface does not change API-level permissions.
Instructions hidden in documents
Prompt injection occurs when untrusted content, such as a website or document, attempts to change model behavior. This is distinct from an ordinary factual error. System instructions must be separated from externally sourced data.
Securing AI integrations requires restricting available actions and validating operation parameters outside the model. Processing a document must not grant new permissions.
Least privilege and approvals
An agent that reads a calendar carries different risks from one that can delete entries or forward files. OWASP describes excessive agency as LLM06:2025. Permissions should be kept to the minimum necessary, with read access separated from changes to data.
When consulting on AI security, determine which actions may run automatically and which require independent confirmation by the user.
Verification and incident response
NIST AI RMF 1.0 describes an iterative process of identifying, measuring and managing risk. In production, this also requires the ability to investigate a failure, limit its effects and restore stable operations.
Configuration histories, operation identifiers and regression tests help establish whether an incident arose from data, instructions, the model or a connected tool.
Sources
- OWASP Gen AI Security Project, “OWASP Top 10 for Large Language Model Applications”, 2025 edition, categories LLM01, LLM02 and LLM06.
- Elham Tabassi, “Artificial Intelligence Risk Management Framework (AI RMF 1.0)”, NIST AI 100-1, 26 January 2023, chapter 5.