Agentic AI in Banking: An OWASP LLM06 Tool-Permission Checklist for UAE Banks

Pay attention to this number: 14%. Any honest talk about agentic AI in banking in the UAE should start there. Not with the demos. Not with the pilots. With deployment.

The DFSA's 2026 AI survey covered DFSA authorised firms in the DIFC. "More than half (53%) of respondents reported exploring, testing, piloting, or deploying AI agents, although only 14% reported having reached deployment stage" [7]. Fifty-three exploring. Fourteen deployed. That gap is where the hard questions live.

Here's why. A chatbot talks. An agent ACTS. OWASP describes agency as the ability to call functions or interface with other systems [1]. The DFSA's research paper describes agentic systems that "use connected tools and take actions under delegated authority" [6]. Delegated authority. In a bank. Think about what that means for a payment, a beneficiary, a customer record.

So the question changes. Not "is the answer good?" The question is "what can this thing change, and who stopped it?" The risk sits in the permission, not the prose.

I'm Ercin Dedeoglu, an AI Architect working on LLM, RAG and agentic systems for regulated banking. Below is the tool-permission checklist I want an architecture or security team to run before any agent touches a production system. Twelve items. Every one tied to a source, or clearly marked as my own practice.

What Is Agentic AI in Banking, and Where Does LLM06 Fit?

For this checklist, agentic AI in banking means an AI system that uses connected tools to carry out banking tasks under delegated authority [1][6]. Action through an interface. That is the line. Fluent conversation is not.

OWASP's own words on agency: "the ability to call functions or interface with other systems via extensions (sometimes referred to as tools, skills or plugins by different vendors) to undertake actions in response to a prompt" [1].

And the failure mode. LLM06:2025 Excessive Agency is one entry in the OWASP Top 10 for LLM Applications 2025 [2]. OWASP defines it like this: "Excessive Agency is the vulnerability that enables damaging actions to be performed in response to unexpected, ambiguous or manipulated outputs from an LLM, regardless of what is causing the LLM to malfunction" [1].

Read the last part again. Regardless of the cause. A hallucination, a bad prompt, an attacker. Doesn't matter. The damage comes from what the system was allowed to do.

OWASP says the root cause is "typically one or more of: excessive functionality; excessive permissions; excessive autonomy" [1]. Three excesses:

  • Excessive functionality. The tool can do more than the task needs.
  • Excessive permissions. The tool's identity can reach more than the task needs.
  • Excessive autonomy. High-impact actions go through without independent approval.

The triggers OWASP lists include hallucination, direct and indirect prompt injection, a compromised extension and, in multi-agent systems, a compromised peer agent [1]. Indirect prompt injection is the nasty one for banks. OWASP explains that it occurs "when an LLM accepts input from external sources, such as websites or files", and lists impacts such as "Providing unauthorized access to functions available to the LLM" and "Executing arbitrary commands in connected systems" [3].

BTW, OWASP now also publishes a Top 10 for Agentic Applications for 2026 [4]. This article stays with LLM06, because LLM06 is where the permission controls are spelled out.

What Do UAE Regulators Already Expect From an AI System That Can Act?

Start with the text. Not the headlines.

The CBUAE guidance note on AI and ML, issued on 11 February 2026, does not use the words "agent" or "agentic" [5]. Its definition of AI does include GenAI "which may have varying levels of autonomy" [5]. So applying it to agents is my mapping, not the note's terminology. Here are the clauses I build from:

  • 2(b): LFIs "should not employ AI models that they have no control over" [5].
  • 2(f): an inventory of all AI models, systems or technologies, with metadata "to include as a minimum model name, purpose, risk rating" [5].
  • 5(d): privacy-by-design and security-by-design, safeguards against unauthorised access or misuse, and stress testing and validation so AI does "not produce unsafe or unintended outputs" [5].
  • 6(f): "the clear and immediate ability, with human intervention, to cease use" of an AI system [5].
  • 7(a): three oversight models, where human-out-of-the-loop "should only be utilised for low-risk, non-material processes with appropriate controls in place" [5].
  • 7(b): human involvement "commensurate with the identified and potential risks posed to a consumer" [5].
  • 8(d): a risk rating for each AI system, influenced by factors including the "capability of the AI, controls in place, impact of the AI" [5].

Control. Inventory. Testing. Oversight. A way to stop it. That's the whole shape.

In the DIFC, the DFSA sent a Dear SEO letter on 4 June 2026 setting out its expectations on AI risk management. It is plain about it: "At this stage, we do not propose to introduce new or amended rules" [8]. Existing rules apply. Firms "remain responsible for ensuring that the use of AI is appropriately governed and controlled" [8]. On cyber risk: "These expectations apply equally to cyber risks arising from a firm's use of AI" [8]. And relying on external model providers, data vendors, cloud service providers or software suppliers does not move that responsibility anywhere [8].

No new rulebook. No excuse either.

The DFSA paper with C:>DIR adds a useful lens, as research framing rather than a DFSA requirement (the DFSA says so itself) [6]. Two of its four dimensions read like an agent design review. Authority: "whether consequential actions by AI-enabled systems remain within their intended mandates". Trust: "whether institutions can establish what happened, bring unfinished activity under control, and continue or restore critical operations reliably" [6].

Mandates and evidence. Exactly what a tool-permission review checks.

I mapped the CBUAE note at board level in my CBUAE AI guidance note article. That piece is governance. This one is engineering: scopes, identities, enforcement, logs, the off switch.

Why Is the Tool Permission, Not the Prompt, the Control That Matters?

Because the model should NOT authorise itself.

OWASP's complete mediation mitigation says it in one line: "Implement authorization in downstream systems rather than relying on an LLM to decide if an action is allowed or not" [1].

That is the boundary I design around. A system prompt that says "never transfer more than the limit" is a request. It is not a control. Prompt injection exists precisely because content can steer the model [3]. If the only thing between injected text and a payment is the model's good behaviour, there is nothing between them.

So the limit lives in the receiving system. The model proposes. The core system decides. My rule is short: a persuasive request must never create a permission.

Testing still matters. A lot. My prompt injection testing article for banking chatbots turns untrusted fields into a go-live gate. With agents the question gets sharper: what action can injected content trigger, and what independent control blocks it?

The Tool-Permission Checklist for Agentic AI in Banking

Run this per agent, per tool. Each item says what to check and what it ties to. Where a control goes beyond the source, it is marked as my practice.

  1. Inventory every tool.

    For each agent, record every tool, its owner, purpose, read or write capability and the downstream system it touches. Remove tools left over from trials. OWASP names that exact exposure: an extension "trialled during a development phase and dropped in favor of a better alternative" that "remains available" [1].

    Ties to: OWASP LLM06 mitigation 1, Minimize extensions [1]. The CBUAE 2(f) inventory covers models, systems and technologies [5]. Recording each tool is my practice, extending it.

  2. Expose the minimum functions.

    Check that each tool exposes only what its task needs. OWASP's example is a document-reading extension that can also modify and delete documents [1]. Review the callable interface itself, not the tool's description.

    Ties to: OWASP LLM06 mitigation 2, Minimize extension functionality [1].

  3. Ban open-ended tools.

    No generic shell. No unrestricted URL fetch. No raw SQL tool. Build narrow operations instead. OWASP advises avoiding open-ended extensions "where possible" and using "extensions with more granular functionality" [1].

    Ties to: OWASP LLM06 mitigation 3, Avoid open-ended extensions [1]. The specific bans are my practice.

  4. Scope the tool's identity to least privilege.

    Check the effective permissions of the identity the tool actually uses to connect. OWASP's example is a read extension connecting with "an identity that not only has SELECT permissions, but also UPDATE, INSERT and DELETE permissions" [1]. A read tool gets a read-only role. Period.

    Ties to: OWASP LLM06 mitigation 4, Minimize extension permissions [1].

  5. Separate read from write.

    Put reading and writing in different tools with different identities. OWASP's mailbox scenario shows why: a plugin that can read and send lets an indirect prompt injection forward sensitive data, and the fixes include a read-only extension and a read-only OAuth scope [1].

    Ties to: my practice, extending OWASP LLM06 mitigations 2 and 4 [1].

  6. Delegate the user's authority, never a shared service account.

    Actions taken for a customer or a staff member should run in that person's context with minimum privileges. OWASP warns against an extension that "accesses downstream systems with a generic high-privileged identity" [1]. One shared super-account behind every agent session is the pattern to kill.

    Ties to: OWASP LLM06 mitigation 5, Execute extensions in user's context [1].

  7. Require human approval for irreversible and money-moving actions.

    Payments, beneficiary changes, limit changes, closures: a human approves before execution. Show the proposed action in an approval step outside the model's chat output, so the model cannot describe its way past it. OWASP: "require a human to approve high-impact actions before they are taken" [1]. CBUAE 7(a) keeps human-out-of-the-loop for "low-risk, non-material processes" and 7(b) asks for human involvement commensurate with the risk [5].

    Ties to: OWASP LLM06 mitigation 6, Require user approval [1]. CBUAE 7(a) and 7(b) [5]. Which actions count and how approval is displayed is my practice.

  8. Enforce authorisation in the downstream system.

    Payment limits, beneficiary rules and entitlement checks run in the core system, on every request, whatever the agent says. A model saying "approval exists" is not approval.

    Ties to: OWASP LLM06 mitigation 7, Complete mediation [1].

  9. Set rate and amount limits.

    Limit actions per user, per tool and per time window. Cap amounts for anything that moves value. OWASP recommends rate limiting "to reduce the number of undesirable actions that can take place within a given time period", and is clear that this limits damage rather than preventing Excessive Agency [1].

    Ties to: OWASP LLM06 damage limitation, rate limiting [1]. The amount caps and the per-user, per-tool split are my practice.

  10. Log every tool call.

    Record who asked, which agent, which tool, the inputs, the output, the approval and the result. EVERY call. OWASP recommends logging and monitoring "the activity of LLM extensions and downstream systems" [1]. CBUAE 5(a) asks for "clear provenance and audit trails" for the data AI models use [5]. The DFSA paper's Trust dimension, establishing "what happened", is the research framing for why this matters after an incident [6].

    Ties to: OWASP LLM06 log and monitor [1]. Extending CBUAE 5(a) audit trails from data to tool calls, and the field list, are my practice.

  11. Build and rehearse a kill switch.

    One switch per agent and one per tool, tested with a human operator before launch and on a schedule after. CBUAE 6(f) asks for "the clear and immediate ability, with human intervention, to cease use" [5]. A provider-independent LLM gateway is a natural place for it.

    Ties to: CBUAE 6(f) [5]. Per-agent and per-tool switches and rehearsal are my practice.

  12. Red-team prompt injection that leads to tool misuse.

    Before launch and on every change, test whether injected content in emails, documents, transaction fields or web pages can push the agent into an unauthorised tool call. Include provider updates. CBUAE 5(d) asks for stress testing and validation, and 6(c) asks that "automatic updates their AI tools are tested before implementation" [5]. OWASP LLM01 mitigation 4 says "Enforce privilege control and least privilege access" [3]. My LLM red teaming checklist for banking covers the scoring and go or no-go gates.

    Ties to: CBUAE 5(d) and 6(c) [5]. OWASP LLM01 [3]. The every-change trigger is my practice.

Twelve lines. Print them. Run them.

How Do You Run the Checklist Without Slowing Delivery?

Make it one table, not a committee.

Folks, the trick is boring. One row per tool. Each row holds the owner, purpose, permitted operations, executing identity, downstream enforcement, approval rule, limits, log location and stop control. When the row is complete, the tool is reviewable. When it isn't, it doesn't ship.

Then split ownership so nobody waits on anybody:

TeamWhat it owns
ArchitectureTool inventory, interfaces, identities and permission scopes
SecurityPrompt injection and tool misuse tests, log checks, evidence review
BusinessApproval thresholds and which actions need a human decision

The pre-launch gate review looks at real configuration and real test evidence. I want to see the restricted identity, a blocked unauthorised request, an approval record and a working stop. A diagram doesn't answer any of those.

Re-run the row when a tool, scope or provider version changes. A new tool raises the capability of the system, and capability is one of the factors CBUAE 8(d) lists for the risk rating [5]. Re-rate it.

Hosting and permissions are separate reviews. Both matter. My sovereign AI architecture for UAE banking covers where the model runs and where data stays. This checklist covers what the agent is allowed to do. One doesn't replace the other.

So Which Approach Should You Pick?

Start read-only. Then earn every write.

Read-only is not risk-free. Sensitive Information Disclosure has its own entry in the OWASP list [2]. But a read-only agent gives your team a small, inspectable permission set to get right first.

Then add write tools one at a time. Each one arrives with a defined purpose, a scoped identity, downstream enforcement, an approval rule where money or irreversibility is involved, limits, logs and a rehearsed stop. No bundle releases. No "we'll add the approval step later".

My recommendation for agentic AI in banking is simple. Grow capability only as fast as you can inspect and control the permission behind it. And NEVER give an agent a permission you cannot log and switch off.

FAQ

What is agentic AI in banking?

It is an AI system that uses connected tools to carry out banking tasks under delegated authority. That working definition draws on OWASP's description of agency and the DFSA paper's description of agentic systems [1][6]. The difference from a chatbot: it can request actions, not just produce text.

Does the CBUAE regulate agentic AI?

The CBUAE note does not use the words "agent" or "agentic" [5]. Its AI definition covers GenAI with "varying levels of autonomy", and it applies to licensed financial institutions using AI [5]. My reading is that its clauses on control, inventory, testing, human oversight and the ability to cease use apply to agents too. The agent-specific mapping in this article is mine.

Is OWASP LLM06 a regulatory requirement in the UAE?

No. OWASP LLM06 is an industry security reference published by the OWASP Gen AI Security Project [1][2]. I use its mitigations as the engineering detail under the CBUAE and DFSA text cited here. The mapping is mine, not a regulator's checklist.

More on my work: About Ercin Dedeoglu.

Sources

  1. OWASP Gen AI Security Project. LLM06:2025 Excessive Agency. https://genai.owasp.org/llmrisk/llm062025-excessive-agency/

  2. OWASP Gen AI Security Project. OWASP Top 10 for LLM Applications 2025. https://genai.owasp.org/llm-top-10/

  3. OWASP Gen AI Security Project. LLM01:2025 Prompt Injection. https://genai.owasp.org/llmrisk/llm01-prompt-injection/

  4. OWASP Gen AI Security Project. OWASP Top 10 for Agentic Applications for 2026. https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/

  5. Central Bank of the UAE. Guidance Note on the Consumer Protection and Responsible Adoption and Use of Artificial Intelligence and Machine Learning by Licensed Financial Institutions in the U.A.E., issued 11 Feb 2026 (sections 1, 2(b), 2(f), 5(a), 5(d), 6(c), 6(f), 7(a), 7(b), 8(d)). https://rulebook.centralbank.ae/en/rulebook/guidance-note-consumer-protection-and-responsible-adoption-and-use-artificial-intelligence

  6. Dubai Financial Services Authority. DFSA and C:>DIR publish new paper on cyber resilience implications of frontier and agentic AI, 2 Nov 2026. https://www.dfsa.ae/news/dfsa-and-cdir-publish-new-paper-cyber-resilience-implications-frontier-and-agentic-ai

  7. Dubai Financial Services Authority. New DFSA AI Survey 2026 findings: AI adoption among DIFC financial services firms has more than doubled since 2024, 2 Nov 2026. https://www.dfsa.ae/news/ai-adoption-among-difc-financial-services-firms-has-more-doubled-2024

  8. Dubai Financial Services Authority. Dear SEO letter: DFSA regulatory expectations on artificial intelligence risk management in the DIFC, 4 Jun 2026 (linked from https://www.dfsa.ae/your-resources/publications-reports/seo-letters-1). https://365343652932-web-server-storage.s3.eu-west-2.amazonaws.com/files/4517/8097/9123/20260604_Dear_SEO_letter_-_Artifical_Intelligence.pdf