Learn
How to keep AI agents from running the wrong tool call
Sourced explainers on checking agent actions before they run: what a pre-execution guard is, the threats it covers, and how to set policies and thresholds you can defend.
Checks run on our GPUs in Switzerland. Request payloads are not stored.Security
8 pages · press / to search
Start here
- ExplainerFoundationsWhat is a pre-execution guard for AI agents?A pre-execution guard checks an AI agent’s tool call before it runs and returns allow, ask or block. What it sees, what it returns and where it fits.Read
- ExplainerThreatsPrompt injection through tool resultsHow instructions hidden in web pages, emails, issues and files hijack agents through tool results, and how to check the resulting actions before they run.Read
- ExplainerSandboxed agentsGuarding code-mode and sandboxed agentsIn code mode an agent writes code that calls many tools at once. Why every call should pass the sandbox’s tool proxy, and how to check them without slowing runs.Read
Deep dives
- Deep divePolicyBlast radius and reversibility: how much can this tool call break?How to reason about the impact of an AI agent tool call: read-only, reversible writes and production side effects, and why backups and dry runs change the verdict.Read
- Deep divePolicyHuman-in-the-loop approval policies that people do not learn to ignoreApproval prompts on every tool call get rubber-stamped. How to route only uncertain or high-impact agent actions to a human, and how to tell if your policy works.Read
- Deep diveThreatsHallucinated tool arguments: the right tool with the wrong valuesAgents invent ids, amounts, recipients and file paths that nobody asked for. How to catch ungrounded tool arguments before the call runs, and what to send the guard.Read
- Deep diveEvaluationCalibrated scores and thresholds: why per-deployment calibration mattersWhat calibrated risk scores mean, why a threshold is a cost trade-off, and why thresholds fitted on one distribution may not hold on yours. A practical recipe.Read
- Deep diveApproachesGuard models vs LLM judges: the cost and latency trade-offSmall guard model or large LLM judge for agent tool calls? The cost, latency and accuracy trade-offs, when a judge is still better, and how to cascade them.Read