ManySpace
Türkçe Sign in Sign up

AI & MLRecommendedAIML-AGT-010

Agent Safety: Permissions, Sandboxing, and Human Oversight

Limiting what a tool-using agent can do: least privilege, sandboxes, confirmation gates, and audit trails.

Difficulty
3 / 5
Importance
4 / 5
Study time
1–2 h
Kind of work
Mixed

Why it matters

The learner can design a permission policy for a tool-using agent that separates reversible from irreversible actions and name the threat each control addresses.

The route to this topic

Route AIML-AGT-010
What Is a Model? Rational Agents and Task Environments Learning from Data Tasks: Regression, Classification, and Beyond Features and Targets Datasets and Splits Classification Framing Linear Regression I: Prediction and Residuals Embeddings Tokens and Tokenization Sampling and Splits Train, Validation, and Test Errors Scores and Logits Loss and Fitting: Squared Error Similarity in Vector Space Distribution Shift Offline Evaluation Multiclass Scores and Softmax Autoregression Objective Functions and Empirical Risk Query, Key, Value Robustness Evaluation Next-Token Prediction Training vs Serving Similarity Scores Adversarial Examples The Agent Loop Weighted Aggregation Tools and Tool Use Threat Modeling for ML Systems Self-Attention Context and Conditioning Prompting and In-Context Learning Prompt Injection Agent Safety: Permissions, Sandboxing, and Human Oversight