Search for a command to run...
Articles tagged with #llm
MCP has four annotations that let a client know whether a tool reads, mutates, deletes, or reaches the open world. Almost nobody sets them - so the agent guesses. hintgap is a static scan that fails the build when a tool ships without them.
LLM-as-judge has quietly become how teams score their AI - but when the same model writes an answer and grades it, self-preference bias inflates the number and you ship a regression on a score you can't trust. Here is why biased judges are an eval bug, and judgeskew, a static gate that fails the build before that score reaches a dashboard.
A retrieval-augmented answer can be fluent, well-structured, and mostly right - and still smuggle in one fabricated fact the retrieved context never supported. Here is why groundedness is the RAG failure mode that actually bites, and nogrounds, a checker that breaks an answer into atomic claims and flags the ones your context can't back up - before the answer reaches a user.
In a multi-tenant RAG app every customer's embeddings usually share one index, and similarity search returns the globally nearest chunks - not yours. Leave the tenant filter off and the search quietly serves someone else's data, passing every single-tenant test. Here is why it leaks, the one filter that fixes it, and ragmoat, a linter that fails the build when the filter is missing.
The model suggests a shell command and your assistant runs it; the model returns some markdown and your page renders it as HTML. Feed it the right input and it emits `rm -rf ~` or `<script>steal()</script>` - and your app obediently does it. This is OWASP LLM02, and here is askance, a static linter that fails the build when model output flows into a dangerous sink.
