LLM security notes

What Are UAB GenAI Expectations?

UAB expectations around generative AI are available on the UAB AI Website. All UAB students, staff, faculty, and collaborators must remain aware of these expectations during the course of their UAB-related duties, when using UAB resources, and when working with UAB sensitive and restricted data. The UAB AI - Responsible AI Use Page has specific information about principles, guidance, and policies.

Research Computing adheres to UAB expectations.

Is My GenAI Tool Approved?

The UAB AI - Tools and Projects Page has a complete list of approved generative AI tools. Tools not on the list will need to be approved on a case-by-case basis. To seek approval for your generative AI use case, visit and complete the AI Request Form at ServiceNow.

What Risks Does Agentic GenAI Pose?

There is an unavoidable tradeoff between autonomy and security. To function autonomously, agentic LLMs need read-write permissions and network access. With both they can do all of the following.

• Permanent deletion of critical data (fs-rw)
• Permanent deletion of data backups (fs-rw, nw)
• Exfiltration of sensitive and restricted data (fs-r, nw)
• Force-push deletion (unrecoverable) of code repositories (fs-rw, vc-rw, nw)
• Exposure of secrets such as passwords, tokens, and keys.
  • (fs-r) if stored in plaintext or command history
  • (env) if stored in environment variables
• Legal liability from exposure (fs-r, nw)
  • Copyrighted material (e.g., journal articles)
  • FERPA-protected data (e.g., student coursework, blazerid)

Relevant permissions are shown in parentheses with shorthand. Their meanings are given below.

• fs-rw is filesystem read-write access.
• nw is network or internet access.
• vc-rw is version control read-write access.
• env is environment variable access.

Use the issue body as the source for the LLM security notes and first locate the appropriate documentation page or section in the repository. Done means the UAB links, risk descriptions, and permission shorthand definitions are published accurately and render correctly.