Why an AI Agent Needs More Than One Safety Check
16 entriesAI agents can retrieve data, call APIs, execute tasks and make decisions, so without proper guardrails they can expose sensitive information or perform unauthorized actions. A LinkedIn post by Rathnakumar Udayakumar (31 October 2026) argues that guardrails are not a single filter but checkpoints throughout an agent's workflow, and names five places to put them: input guardrails that stop unsafe requests before they reach the agent, retrieval guardrails that keep the agent to trusted information, tool-call guardrails that control what it can execute, output guardrails that validate responses before delivery, and human approval gates for high-risk actions. This page follows those five checkpoints, lists the tools the post names with others, and adds testing and standards resources. Descriptions are paraphrased from each source, are not rankings, and no claim is made that any tool prevents all attacks. The post's immigration-services promotion is not part of this page.
1. Input Guardrails
7 entriesStop unsafe requests before they reach the agent: detect prompt injections and jailbreaks, redact sensitive information and filter malicious inputs.
2. Retrieval Guardrails
5 entriesEnsure agents access trusted information: enforce document permissions, detect malicious retrieved content and filter irrelevant sources.
3. Tool-Call Guardrails
8 entriesControl what agents can execute: allowlist approved tools, validate API arguments, set permission and spending limits.
4. Output Guardrails
6 entriesValidate responses before delivery: check factual grounding, prevent sensitive data leaks and validate output schemas.
5. Human Approval Gates
5 entriesKeep humans involved in high-risk actions such as large refunds, sensitive data deletion and customer-facing actions.
Testing, Standards and Threat Models
8 entriesWays to test guardrails and frameworks that define the risks they address.
Frequently Asked Questions
What are guardrails for an AI agent?
Checks and limits placed around an agent so that unsafe requests, untrusted data, risky tool calls and unvalidated answers are caught before they cause harm. The LinkedIn post this page follows says guardrails are checkpoints throughout the workflow, not a single filter.
Where should guardrails go?
The post names five places: on inputs, on retrieval, on tool calls, on outputs, and at human approval gates for high-risk actions.
What do input guardrails do?
They stop unsafe requests before they reach the agent by detecting prompt injection and jailbreaks, redacting sensitive information and filtering malicious input, using tools such as Llama Guard and Presidio.
What are retrieval guardrails?
Controls that keep the agent to trusted information: enforcing document permissions, detecting malicious retrieved content and filtering irrelevant sources, because retrieved text can carry injected instructions.
How do I guard tool calls?
Allowlist approved tools, validate API arguments, set permission and spending limits, and run risky actions in sandboxes. A policy engine such as Open Policy Agent can enforce rules outside the model. See the AI Agent Security Solutions page.
When should a human approve an action?
For high-risk or irreversible actions, for example large refunds, sensitive data deletion or customer-facing actions, as the post suggests. LangGraph's interrupt feature can pause a workflow for approval.
Do guardrails make an agent safe?
No. Guardrails reduce risk but can be bypassed, so combine them with least-privilege access, monitoring, testing with tools such as garak and promptfoo, and governance. See the AI Governance Frameworks page.
How current is this page?
Links and release facts were checked on 1 November 2026 from GitHub. LinkedIn may ask visitors to sign in. Descriptions follow each source's own pages and are not rankings or guarantees.