Bengaluru Tech Summit 2026 · 17–19 Nov · BIEC · Hall 1 · Booth H1A106
Control what your AI agents can see and do - inside your own cloud.
DeepintShield is the self-hosted AI security and agent control plane for banks, payment companies and global capability centres. Policy on every prompt, retrieval, tool call and output. Approvals for high-risk actions. A full audit trail. Nothing leaves your VPC.
Mapped to the OWASP LLM Top 10, the OWASP Agentic Top 10 and NIST AI RMF.
What goes wrong
Your board will ask how you control AI agents. Bring the answer to BTS.
Live at the booth
What we will show at booth H1A106
Bring one prompt you would like us to attack. Each demo takes about four minutes.
Prompt injection, blocked.
A poisoned document tries to make an agent leak customer data. Watch the input- and retrieval-stage policy stop it - and log it.
Tool call, held for approval.
The agent tries to send a payment instruction. The REQUIRE-APPROVAL decision pauses it until a human says yes.
MCP call, denied.
An unregistered MCP server is called. Denied, logged, alerted.
Free checklist
25 controls your auditor will ask about - in one checklist
Governance, input, retrieval, action, output and monitoring controls for GenAI and agent deployments, each with the evidence to show and a self-score. Written for CISOs, risk teams and AI platform leads.
- Governance and inventory (5 controls)
- Input stage: prompts and users (5)
- Retrieval stage: RAG and knowledge (3)
- Action stage: tools and agents (5)
- Output stage and monitoring (7)
Mapped to the OWASP LLM Top 10, the OWASP Agentic Top 10 and NIST AI RMF.
Thank you - your checklist is below.
We will also email you a copy. Want us to score your real deployment? Book 20 minutes at the booth or on a call.
One more step for your seat: confirm it on Luma, which sends you the join link and the reminders.
The full checklist opens here as soon as you send the form - all 25 controls with the evidence to show and how to score yourself.
25 controls for GenAI and agent deployments
A practical map to the OWASP LLM Top 10, the OWASP Agentic Top 10 and NIST AI RMF; each control comes with the evidence a security review asks for.
Governance and inventory (5)
- 1Board-approved AI/model risk policy covering GenAI and agentsEvidence: Policy document, approval minutesHow DeepintShield helps: Policy-as-code export that mirrors the written policy
- 2Complete inventory of AI applications, models and providers in useEvidence: Model and app register with ownersHow DeepintShield helps: Automatic register of every app, model and provider seen at the gateway
- 3Named owner for each AI application and a risk tierEvidence: Register with owner and tierHow DeepintShield helps: Tier field drives policy strictness per app
- 4Third-party and open-source model risk assessed (provenance, licence, update path)Evidence: Vendor/model assessment formsHow DeepintShield helps: Provider-neutral routing lets you switch models without re-engineering
- 5Human-in-the-loop defined for high-risk decisions and actionsEvidence: Decision matrix: what needs approvalHow DeepintShield helps: REQUIRE-APPROVAL decision type with approver roles
Input stage (prompts and users) (5)
- 6Prompt-injection and jailbreak detection on every inputEvidence: Test results; blocked-attempt logsHow DeepintShield helps: Input-stage policies; attack-pattern detection; logs
- 7Identity and authorization for every caller (user, service, agent)Evidence: Auth design; access matrixHow DeepintShield helps: ReBAC via OpenFGA per user, team and agent
- 8Masking of personal and sensitive data before it reaches a modelEvidence: DPDP data map; masking rulesHow DeepintShield helps: MASK decision with configurable detectors
- 9Rate limits and abuse controls per user and per appEvidence: Limits config; alertsHow DeepintShield helps: Rate-limit-aware load balancer and per-app quotas
- 10Separation of system prompts and customer data; no secrets in promptsEvidence: Prompt templates reviewHow DeepintShield helps: Policy checks on prompt templates and secret patterns
Retrieval stage (RAG and knowledge) (3)
- 11Document-level access control on retrieval (users only retrieve what they may read)Evidence: Index permissions; test casesHow DeepintShield helps: Retrieval-stage ReBAC filter
- 12Indirect prompt-injection defence for retrieved contentEvidence: Tests with poisoned documentsHow DeepintShield helps: Retrieval-stage scanning and quarantine
- 13Data residency: knowledge and embeddings stay in approved regionsEvidence: Architecture diagram; cloud configHow DeepintShield helps: Self-hosted; zero data egress by design
Action stage (tools and agents) (5)
- 14Allow-list of tools and actions per agent; everything else deniedEvidence: Tool catalogue with policiesHow DeepintShield helps: Action-stage policies per agent and tool
- 15Approval required for money movement, customer-data changes, external sendsEvidence: Approval logs with approver identityHow DeepintShield helps: REQUIRE-APPROVAL workflow with audit
- 16Kill-switch: any agent or app can be suspended immediatelyEvidence: Runbook; drill recordHow DeepintShield helps: One-click DENY-ALL policy per app or agent
- 17MCP servers registered, scoped and monitored; no unknown MCP endpointsEvidence: MCP register; policiesHow DeepintShield helps: MCP-stage enforcement and discovery
- 18Least privilege for agent credentials; no shared admin keysEvidence: Credential inventoryHow DeepintShield helps: BYOK and scoped provider keys
Output stage and monitoring (7)
- 19Output filtering for sensitive data, harmful content and hallucinated commitmentsEvidence: Output policies; samplesHow DeepintShield helps: Output-stage MASK/DENY decisions
- 20Complete, tamper-evident audit trail of prompts, retrievals, actions, decisionsEvidence: Audit export reviewed by internal auditHow DeepintShield helps: Full audit trail; export to SIEM
- 21Continuous monitoring and alerting for AI-enabled attack patternsEvidence: SOC runbook; alert examplesHow DeepintShield helps: Decision telemetry to the SOC/SIEM
- 22Security testing before go-live and after material change (red-team, regression)Evidence: Test reports; change logHow DeepintShield helps: Red-team playbook and regression suite in the Assessment
- 23Incident response playbook for AI incidents (prompt leak, tool misuse, model outage)Evidence: IR plan; tabletop recordHow DeepintShield helps: Decision logs feed IR; kill-switch drill
- 24Consistency and drift checks on model responses across providersEvidence: Evaluation resultsHow DeepintShield helps: Response Consistency Engine
- 25Periodic review and reporting to the board (metrics, incidents, exceptions)Evidence: Quarterly reportHow DeepintShield helps: Dashboard and exportable reports
Live session · Thursday 12 November · 4:00–4:40 pm IST
Control the agent: securing AI agents in production
What goes wrong when AI agents go live, the 25 controls that stop it, two live attacks on an agent and the fix, and your questions. Built for CISOs, security and risk teams and AI platform leads in banks, NBFCs, payment companies, AI-native companies and GCCs.
- 40 minutes, online
- Two live attacks on an AI agent, and the fix
- The three controls auditors ask about first
- Your questions, answered live
Start small
Two ways to start
AI Agent Security Readiness Assessment
- Red-team of 1–3 of your GenAI or agent apps
- Control-gap map against the 25 controls
- DeepintShield deployed observe-only in your cluster
- Board-ready report and read-out
- Fee credited against your subscription
BTS Pilot
- Full enforcement for one business unit
- Written success criteria
- Weekly check-ins
- Fee credited against the annual plan
- Limited to eight pilots starting before 31 December
Visiting BTS 2026
Find us at Hall 1, booth H1A106
17 Nov from 9:00 · 18–19 Nov from 10:00
Open in Google Maps
18–19 Nov: free Business Visitor registration
BTS registration
Book 20 minutes
Book 20 minutes with our founder
Pick a slot at booth H1A106 on 18 or 19 November, or a call before or after the event. You get a calendar invite straight away.
Calendar not showing? Open the booking page
Built for regulated enterprises
Runs where your data already lives
FAQ
Questions we hear most
Does any data leave our environment?
No. The control plane and data plane run in your Kubernetes; model calls go to the providers you already use, through your keys.
How long does deployment take?
A Helm install in an afternoon; policies for a first app in a day.
Which frameworks are supported?
LangChain, LangGraph and any OpenAI-compatible client through the gateway; MCP servers are supported natively.
Is this only for banks?
No. GCCs, insurers, AMCs, fintechs and AI product companies use the same controls; the checklist is written for regulated environments.
What does it cost?
The Readiness Assessment and the Pilot are fixed-fee, and the fee is credited against your subscription. Subscription plans are on our pricing page.
Do I need a pass to visit the booth?
17 November is for delegate and premium pass holders. On 18 and 19 November the free Business Visitor registration gets you in. Book a slot and we will hold the time for you.
Can we meet outside BTS?
Yes. Book 20 minutes for a call before or after the event, or email us.
See you at booth H1A106 - bring one prompt you would like us to attack.
DeepintShield · self-hosted AI security for regulated enterprises · support@deepintshield.com