Tag
11 articles tagged LLM.
Privacy
A model aggregator sees every prompt, upload, and output you send through it. The 302.ai exposure shows why that makes it a data processor to vet, not just an API to call.
AI Security
An agent that remembers can be taught to remember the wrong things. Treat every memory write as a privileged action with provenance, review, and an expiry date.
AI Security
The LiteLLM 1.107.0 release on PyPI and its Docker images carried a credential stealer that runs on every Python start. Here is what it did, how to tell if you were hit, and what to rotate.
AI Security
Anything a prompt can influence, an attacker can influence. Here is how to handle model output safely at each place it lands: the browser, the database, the shell, and the interpreter.
AI Security
A system prompt is configuration the user can probably read, not a vault. Here is what to move out of it, where to put it instead, and how to test that it stays out.
AI Security
An LLM endpoint turns every request into a bill. Here is how to cap what a single user, an agent loop, or a stolen API key can spend before it becomes an incident.
AI Security
A RAG system that searches one shared index for every user will eventually hand someone a document they were never allowed to read. Here is how to enforce access control at retrieval time.
AI Security
AI coding assistants regularly suggest packages that do not exist, and the same fake names come back again and again. Here is how slopsquatting works and the controls that stop it at review, install, and CI time.
AI Security
A downloaded model checkpoint is not just data. If it is a pickle, loading it can execute code. Here is how to load third-party weights without handing that file a shell.
AI Security
Filters and stern system prompts reduce prompt injection, but they cannot eliminate it. The durable fix is to treat model output as untrusted input and authorize every action it proposes.
Privacy
LLM applications need detailed logs to debug and investigate incidents, but prompts are full of personal data. Splitting logs by purpose and redacting at the edge resolves most of the tension.
