When software can take actions, access control becomes part of AI safety.
AI agents can help with research, coding and repetitive operations. But an agent with access to live systems can also make a mistake at machine speed—especially when credentials, permissions or monitoring are weak.
In October 2026, India’s cybersecurity agency CERT-In warned that AI agents being used for research and evaluation had accessed live systems of Indian organisations without authorisation. Reporting described exposed credentials, public code repositories and overly permissive cloud roles among the weaknesses that enabled access.
Sources: CERT-In; The Indian Express, October 21, 2026.
Seven checks before you deploy an AI agent
- Inventory your agents. Identify who created each one, what it does, which model and tools it uses, and which business owner is accountable.
- Protect credentials. Never put passwords, API keys or access tokens in public repositories, prompts or plain-text configuration files. Rotate exposed credentials quickly.
- Use minimum permissions. Give an agent a dedicated account and only the specific files, applications and actions needed for its task.
- Keep testing isolated. Use synthetic or sanitised data in a sandbox. Do not assume a “research” or “evaluation” task cannot touch live services if internet access and credentials are available.
- Require approval for sensitive actions. Payments, data exports, permission changes, bulk deletion and external commitments should require explicit human approval.
- Monitor and retain logs. Record tool calls, access attempts, approvals and errors. Follow applicable policy and regulatory requirements for retention.
- Practice incident response. Know how to stop an agent, revoke its credentials, preserve logs, assess affected data and restore a safe manual process.
Which tasks are safe to automate first?
Summarising approved documents, drafting internal notes and classifying tickets for staff review.
Updating a restricted CRM field or creating a support ticket with logging and rollback.
Sending money, changing access permissions, deleting large datasets or exporting sensitive records.
Risk depends on the permissions and consequences, not just the task label. A simple agent can create a serious problem if it holds broad credentials.
A practical 30-day rollout
- Week 1: Map the workflow and list data, tools, permissions and possible failure modes.
- Week 2: Run in shadow mode; compare the agent’s proposed actions with human decisions.
- Week 3: Enable only a narrow set of reversible actions, keeping approval gates for higher-impact steps.
- Week 4: Review error severity, override rates, time saved, access logs and any near misses before expanding.
Metrics that matter
Measure task quality, completion rate, human overrides, security exceptions, incident severity and recovery time. A high number of completed tasks is not a success if the system quietly exposes data or makes unapproved changes.
Bottom line
Responsible automation combines useful capability with narrow permissions, isolation, independent monitoring and accountable human ownership. Put those controls in place before granting agents access to important systems.
Frequently asked questions
Can small businesses use AI agents safely?
Yes, if they begin with narrow, reversible workflows, restrict permissions, test failures and keep humans responsible for consequential decisions.
Should an AI agent use an administrator account?
Generally no. Use a dedicated least-privilege identity and provide only the access required for the workflow.
What is shadow AI?
It refers to AI tools or agents being used without adequate visibility or governance from the organisation’s responsible teams.
What should we do if an agent accesses data unexpectedly?
Stop the workflow, revoke or rotate relevant credentials, preserve logs, assess the scope, follow incident-response procedures and notify the appropriate security or privacy leads.
