The Signal: AI agents are beginning to do more than draft answers. They can retrieve records, call APIs, update business systems and coordinate multi-step workflows. That makes identity and authorization a frontline AI-security issue. The risk is not that organisations automate coordination; it is that they allow software to act without a clearly scoped identity, enforceable permissions, accountable owner and reliable audit trail.
- Automation changes the control surface
- The identity gap: which agent did what, with whose authority?
- Why least privilege must be enforced outside the model
- The controls that make autonomous workflows governable
- Indian compliance is a design requirement, not a prompt
- A staged path from pilot to production
- FutureIsNow View : automate the work, preserve the controls
Automation changes the control surface
Traditional business processes often rely on a mix of software permissions and human checks. A manager may question an unusual payment, ask why a customer record is being exported or stop a request that does not fit the business context. Those checks are imperfect and inconsistent, but removing a human step does not remove the need for control.
When an AI workflow can act across email, customer systems, code repositories or finance applications, the organisation must encode the relevant boundaries into the system itself. A prompt that says “be careful” is not an access-control mechanism. Permissions must be enforced by the tools and services the agent calls, even when the model is mistaken, manipulated by untrusted content or pursuing a task in an unexpected way.
This is a growing operational concern, not just a theoretical one. On October 16, 2026, Reuters reported that India’s cybersecurity agency CERT-In had disclosed incidents in which AI systems being tested accessed live business systems and records beyond their intended scope. The incidents underline why test environments, explicit authorization and production boundaries must be designed—not assumed.
Source: Reuters’ report on AI agents accessing live Indian business systems.
The identity gap: which agent did what, with whose authority?
Enterprises already manage non-human identities: service accounts, API clients, automation bots, workload identities and application credentials. Agents add complexity because they can select tools dynamically, interpret untrusted inputs, delegate subtasks and operate across several systems. A single shared credential can make it difficult to establish which agent initiated an action or revoke one workflow without breaking others.
Every production agent should have a named owner and a distinct, attributable identity. Credentials should be short-lived where practical, scoped to a task and resource, and revocable independently. Read and write access should be separated. A workflow that only summarises customer records should not possess permission to modify those records simply because a future version might need it.
The US National Institute of Standards and Technology is actively examining identity and authorization for software and AI agents, including how they are identified, authenticated and granted access to data and tools. This work reflects a practical lesson: agent identity must fit into existing enterprise identity and access-management systems, rather than becoming an untracked parallel layer.
Source: NIST’s concept paper on software and AI agent identity and authorization.
Why least privilege must be enforced outside the model
Agent security breaks down when the model itself is trusted to decide whether an action is permitted. An agent can misinterpret a request, follow malicious instructions hidden in a document, choose the wrong tool or chain several individually valid actions into a harmful outcome. The control layer must independently check the user, agent, operation, target resource and current authorization before executing each consequential action.
OWASP’s 2026 Top 10 for Agentic Applications highlights risks such as excessive agency, tool misuse, identity and privilege problems, and failures in oversight. Its guidance is not a substitute for an organisation’s threat model, but it offers a useful baseline for designing and testing agentic systems.
Source: OWASP Top 10 for Agentic Applications 2026.
The controls that make autonomous workflows governable
- Inventory and ownership: maintain a register of agents, their business purpose, owner, data access, connected tools and deployment environment.
- Unique identities: avoid shared human credentials. Give each agent or workflow a traceable identity and make it possible to disable it independently.
- Least privilege: start with read-only access where possible; allow only the tools, records and operations required for the task.
- Policy enforcement at execution: validate every tool call at the API, gateway or application layer. Do not rely on instructions in the prompt to enforce authorization.
- Human approval for consequential actions: require an explicit approval or a well-defined policy gate for payments, permission changes, bulk exports, destructive actions and production deployments.
- Boundaries around untrusted input: treat web pages, emails, tickets and documents as data—not as trusted instructions. Limit outbound network access and access to secrets.
- Audit trails: log the initiating user, agent identity, requested action, effective permissions, tool result and approval state while protecting sensitive content.
- Operational limits: set budgets, rate limits, retry caps and stop conditions to contain loops, accidental bulk actions and unexpected cost.
- Testing and revocation: test prompt injection, privilege escalation and failure scenarios in isolated environments; review permissions after changes and revoke credentials promptly when needed.
Indian compliance is a design requirement, not a prompt
AI workflows that process personal data or access regulated systems must be assessed against the laws and sectoral obligations that actually apply to the organisation and use case. India’s Digital Personal Data Protection framework does not mean every agent action automatically triggers a violation, nor does it make human review mandatory for every automated step. It does make purpose, safeguards, access, retention and accountability important considerations where personal data is involved.
Security teams should map each workflow to the data it uses, the purpose for which that data is processed, the systems it can change and the evidence needed to investigate an incident. Legal, privacy, security and business owners should agree on the permitted actions before deployment—especially for customer data, employee records, payments and critical infrastructure.
A staged path from pilot to production
- Stage 1 — Observe: run the workflow against test data and record the actions it proposes without allowing state changes.
- Stage 2 — Read-only production: connect to approved live sources with narrowly scoped access; compare outputs against known results.
- Stage 3 — Controlled writes: allow reversible, low-risk changes with validation, rate limits and rollback procedures.
- Stage 4 — Higher-impact actions: introduce explicit approval gates, separation of duties and enhanced monitoring before expanding authority.
- Stage 5 — Continuous assurance: re-test after changes to the model, prompt, tools, data sources or permissions; monitor for unusual behaviour and review access regularly.
This approach connects with the wider enterprise challenges explored in FutureIsNow’s analysis of shadow AI and non-human identity governance, AI agent liability and controlled execution in enterprise vertical AI.
FutureIsNow View : automate the work, preserve the controls
AI agents can reduce coordination overhead and make complex workflows faster. But the business case is incomplete if automation removes friction without replacing the controls that friction sometimes provided. The answer is not to preserve every manual approval; it is to identify which checks protect data, money, safety or business integrity and implement them as explicit, testable policies.
Standards work is beginning to formalise this requirement. In February 2026, NIST’s National Cybersecurity Center of Excellence published a concept paper on applying identity and authorization practices to software agents, explicitly identifying identification, authorization, auditing and non-repudiation as design concerns. The World Economic Forum’s May 2026 playbook, developed with Capgemini, proposes an Agent Capability and Authorization Profile to make delegated authority and actions auditable and enforceable across an agent’s lifecycle. These are guidance and standards-development efforts, not a single mandatory global standard. See NIST’s agent identity concept paper and the WEF playbook on agent authorization.
Bottom line: Treat every agent as a software actor with an identity, owner, limited authority and auditable actions. The goal is not to make agents less useful. It is to make their usefulness safe enough to scale.



