Securing AI Models and Agents
LLM security risks
LLM Security Risks
- The OWASP Top 10 for LLM Applications, Explained
- A Review Checklist for LLM Features Before Launch
- How to Map OWASP LLM Risks to Concrete Controls in Your Stack
- When to Fix the Prompt Instead of Fine-Tuning
- A Cost Table for Prompt Iteration Versus a Training Run
- A Lightweight Threat Model for LLM Features
- Building a Fine-Tuning Set From Your Corrections Log
- Keeping Formatting and Policy Rules Stable Across a Model Swap
- All of LLM security risks
Prompt injection
Prompt Injection
- Prompt Injection: What It Is and How to Defend Against It
- A Severity Rubric for Injection Failures: What Blocks a Release
- How to Test an LLM Application for Prompt Injection
- Indirect Injection Through the Files Your Agent Opens
- A Release Gate That Reads Injection Scores From CI
- A Review Queue for Human-Judged Injection Cases
- Detecting Prompt Injection in Production
- Injection Cases for a Refund Tool
- All of Prompt injection
Agent access control
Agent Access Control
- How to Scope Each Tool an Agent Can Call
- How to Write a Security Review of an AI Agent
- Tool Permissions for LLM Agents: Scoping Access and Data
- Choosing Between a Single Broad Tool and Many Narrow Ones
- Mapping Each Tool to Its Blast Radius
- Testing Tool Permissions Across Agent Roles, Not Just Tenants
- What a Denied Action Should Say to the Model
Logging and monitoring
Logging and Monitoring
- Paging Without a Big Ops Team: On-Call for Small Companies
- A Support-Facing Dashboard for Model Health
- A Trace Format That Survives a Model Swap
- Alerting on What Customers Notice, Not What the Dashboard Shows
- Detecting Model Drift Without Re-running the Full Suite
- Drawing Your Agent Topology From Traces
- Instrumenting Retries: Measuring Repair Rates and Their Cost
- Log Fields That Answer an Offboarding Question Later
- All of Logging and monitoring
Human oversight of agents
Human Oversight of Agents
- What the Agent Says When No Human Answers
- A Pre-Launch Review for a Coding Agent That Opens Pull Requests
- A Tabletop for a Runaway Agent: Walk the Trace Before It Exists
- Counting Rubber Stamps: Measuring the Yes Rate on Agent Approvals
- Designing the Approver Inbox Where One Yes Moves One Item
- Designing the Escalation Ladder Before the Fallback
- Drafting the Internal FAQ for a Model Change
- End User Consent for an Agent That Acts on Their Behalf
- All of Human oversight of agents
Agent testing and red teaming
Agent Testing and Red Teaming
- How to Do a Security Review of an AI Agent
- Adversarial Test Cases for Agents That Use Tools
- Keeping a Security Review Current as Agents Change
- Setting a Budget for Injection Testing Before Launch
- Testing a Model SDK by Its Failure Modes, Not Its Happy Path
- When the Model Picks No Tool: Scoring the Abstain Case
- When to Re-run Evals After a Prompt Edit
- When to Re-run Evals for a Tool Change, Not a Prompt Change
- All of Agent testing and red teaming