Leaked & long-lived secrets
Keys in code, chat threads, notebooks and agent configs that never expire. Impact: silent data access for months.
OWASP NHI2 · NHI7Every AI agent, API key and service account — found, owned, scoped and short-lived.
XcellHost discovers the non-human identities your AI agents, copilots and integrations run on, cuts them to least privilege, moves standing keys to short-lived tokens, and runs owner reviews every quarter — from our Mumbai SOC, billed in rupees.
Every chatbot, copilot, RAG pipeline and AI agent reaches your data through a non-human identity — a service account, an API key, an OAuth app, a managed identity or a newer agent identity. They outnumber employees many times over, they rarely use MFA, and they are often created in a hurry, given broad access "to make it work", and then forgotten. When one leaks, the attacker gets everything that identity could touch.
XcellHost AI Identity Security brings those identities under the same discipline you apply to people. We find them across your clouds, SaaS apps and code, give each one a named owner, cut its permissions to the tasks it actually performs, replace standing secrets with short-lived tokens issued just in time, and review every identity with its owner each quarter. Start with a one-time assessment, or let our SOC run it as a managed service — on a GST invoice in rupees.
Illustrative sample, not survey data. Your real numbers come out of the assessment.
Most AI incidents involving data are, underneath, identity failures: a key that leaked, a scope that was too broad, an account nobody switched off. These are the risk categories we assess — described at category level, with what each one costs the business.
Keys in code, chat threads, notebooks and agent configs that never expire. Impact: silent data access for months.
OWASP NHI2 · NHI7Tenant-wide or write scopes granted "to make it work". Impact: one mistake or misuse reaches every system the agent can.
OWASP NHI5 · ASI03 · LLM06Identities outlive the project, the vendor or the employee who made them. Impact: orphaned access nobody will notice.
OWASP NHI1Manipulated inputs push an agent to use its tools outside the task. Impact: the damage is bounded only by its permissions.
OWASP ASI01 · ASI02 · LLM01Plugins, MCP servers and SaaS integrations holding your tokens. Impact: a supplier breach becomes yours.
OWASP NHI3 · ASI04 · ASI07One key across agents, environments or humans. Impact: no audit trail you can attribute — a DPDP and CERT-In problem.
OWASP NHI8 · NHI9 · NHI10Each agent identity moves through the same loop. Hover a control to see where it sits; the token shows where a live identity is right now.
Every assessment starts with a signed scope and rules of engagement. We work with read-only access wherever possible; any control testing is agreed in advance, non-destructive and run against the environments you name.
Agents, apps, clouds and repos in scope; access method; test windows; contacts; data handling. Signed before work starts.
Week 0Read-only inventory of service accounts, keys, OAuth grants, managed and agent identities, and MCP connections.
Week 1Effective permissions versus actual use, credential age, storage location, sharing and ownership gaps.
Week 1–2Agreed, non-destructive checks that agent tool boundaries, token lifetimes and conditional access behave as designed.
Week 2Findings ranked by business impact, with a least-privilege scope list per agent and a rotation plan.
Week 3Fixed items re-checked; a re-test letter and the option to move into managed governance.
Week 4–5Every finding and control carries its references, so the same work feeds your ISO 42001 programme, your DPDP readiness and your board report.
| Framework | What it covers | Where our work maps |
|---|---|---|
| OWASP Non-Human Identities Top 10 (2025) | Improper offboarding, secret leakage, vulnerable third-party NHI, over-privilege, long-lived secrets, NHI reuse, human use of NHI | Discovery, ownership, secrets & rotation, review — all ten risks |
| OWASP Top 10 for Agentic Applications (2026) | Published Dec 2025; includes identity & privilege abuse (ASI03), tool misuse (ASI02), insecure inter-agent communication (ASI07), rogue agents (ASI10) | Agent scopes, tool allow-lists, agent-to-agent auth, kill-switch & revocation |
| OWASP Top 10 for LLM Applications (2025) | Prompt injection (LLM01), sensitive information disclosure (LLM02), excessive agency (LLM06) | Least privilege limits what a manipulated model can do |
| MITRE ATLAS | Adversary tactics against AI systems, now including agent configuration and credential techniques | Threat model for each agent; detections in our SOC |
| NIST AI RMF 1.0 + GenAI Profile (NIST AI 600-1) | Govern · Map · Measure · Manage | Ownership (Govern), inventory (Map), risk scoring (Measure), remediation (Manage) |
| ISO/IEC 42001:2023 · ISO/IEC 23894 | AI management system; AI risk management guidance | Evidence for AI system inventory, roles, resources & operational controls |
Rules notified in November 2025, with most duties phased in over 18 months. Reasonable security safeguards include access control and logs; a personal-data breach must be intimated without delay, with a detailed report to the Board within 72 hours. Agent identities touching personal data fall inside this.
Reportable incidents within 6 hours and ICT logs kept for 180 days in India. Attributable, per-agent identities make both possible.
The RBI committee's Framework for Responsible and Ethical Enablement of AI: 7 sutras and 26 recommendations for regulated entities, including governance, accountability and security of AI systems.
Proposed guidelines on responsible AI/ML use by market participants — accountability, testing, records and controls. Proposed, not yet final; we track the outcome.
Seven principles including accountability and "safety, resilience & sustainability"; a principles-based approach rather than a new law.
Per CERT-In directions (2022) and DPDP Rules 2025. XcellHost does not provide legal advice.
The EU AI Act's prohibitions applied from 2 February 2025 and general-purpose AI model obligations from 2 August 2025. High-risk obligations were originally due from 2 August 2026; an EU "Digital Omnibus" agreement reached in May 2026 proposes moving them to December 2027 and August 2028 once formally published. Identity, logging and human oversight evidence from our work supports a provider's technical file.
*Proposed dates take effect only on publication in the EU Official Journal. Check current status with your counsel.
Documents your team and auditors can use, plus the target state each agent should run in.
Two quick tools. Both run in your browser; nothing is sent anywhere. Results are illustrative estimates, not an assessment.
Illustrative weighting by XcellHost, not a standard score. The assessment measures your real exposure.
Indicative — confirm with XcellHost. INR, excl. 18% GST. Cloud, vault and licence costs (if any) are billed separately at actuals.
A scoped, one-time engagement with report, risk register and re-test.
Our SOC keeps every agent identity owned, scoped and short-lived.
Multi-cloud, regulated, or hundreds of agents.
Indicative — confirm with XcellHost. *Placeholder prices, scope limits and discounts; INR, excl. 18% GST.
Indicative XcellHost fees. Add to cart, or send it to us for a formal quotation.
3 included; each extra agent or app ₹24,999*.
Agents + API keys + service accounts + OAuth apps. 250 included; ₹7,999* per extra 100.
| Typical AI pilot | Governed by XcellHost | |
|---|---|---|
| Credential | Long-lived API key in a config file | Short-lived token per task, minutes-long TTL |
| Permissions | Broad or admin scope "to make it work" | Only the scopes its tasks need; writes approved |
| Owner | Whoever built the pilot, if anyone | Named owner, purpose and expiry on record |
| Sharing | One key shared by agents, apps and people | One identity per agent and environment |
| Visibility | Calls logged under a generic account | Every tool call attributable to one agent |
| Offboarding | Key keeps working after the project ends | Revoked at decommission; reviewed quarterly |
| Internal self-assessment | XcellHost assessment | |
|---|---|---|
| Coverage | The agents the team remembers | Discovery across cloud, SaaS, code and MCP |
| Mapping | Ad-hoc checklist | OWASP NHI, Agentic, NIST AI RMF, ISO 42001 |
| Prioritisation | By how loud the issue is | By business impact and blast radius |
| Output | A spreadsheet of findings | Report, risk register, scope sheets, re-test letter |
| After the report | Fixes compete with feature work | Optional managed governance by our SOC |
AI Identity Security follows an agreed scope for access, data handling, evidence and follow-up.
Document the AI systems, identities, environments and activities included in the engagement.
Agree the accounts and permissions required for the scoped review or testing.
Agree safe test data and handling rules for prompts, outputs and supporting evidence.
Limit access to collected findings and agree their storage and retention.
Map relevant findings to the frameworks selected for your engagement.
Assign owners to prioritised actions and agree the validation or follow-up scope.
AI identity security, run by a 24×7 SOC in Mumbai.
One tax invoice for assessment and managed service.
01Anomalous NHI use alerted 24×7 by our analysts.
02Microsoft Gold Partner; Entra, cloud IAM and PKI teams in-house.
03ISO 27001 & 20000-1 certified; outputs map to ISO 42001 and DPDP.
04Written rules of engagement; least access for us too.
05Mumbai-based managed cloud and security, phone and WhatsApp support.
06Your agents, clouds and what worries you most.
Signed scope, access and test windows.
Discovery, analysis and findings in 3–5 weeks*.
Scopes cut, secrets rotated, fixes confirmed.
Managed reviews and SOC monitoring, monthly.
Any identity that isn't a person: service accounts, API keys and tokens, OAuth apps, managed identities, workload identities, certificates — and now AI agents. They let software talk to software, and in most companies they far outnumber employees.
Agents act on their own, call many tools and are created quickly by many teams. Each one needs credentials, and each one can be steered by the content it reads. If its identity is over-privileged, a manipulated or faulty agent can do everything that identity allows.
The agent gets only the scopes its approved tasks need — read before write, its own mailbox before all mailboxes — and actions with real-world effect, such as payments or deletions, wait for a human to approve. The permission planner above shows the idea.
Instead of a key that works forever, the agent receives a token when a task starts, scoped to that task and expiring in minutes. If it leaks, it is already useless or soon will be. We design this using your identity provider, cloud workload identity or a vault.
Only within a signed scope and rules of engagement. Identity assessment is mostly read-only analysis; any control validation is agreed in advance, non-destructive and run in named environments. For adversarial testing of the AI application itself, see our AI Red Teaming Services.
Microsoft Entra and Azure, AWS IAM, Google Cloud, common SaaS apps with OAuth, GitHub and GitLab, and AI platforms such as Azure OpenAI, OpenAI, Anthropic, Gemini and AWS Bedrock, plus agent frameworks and MCP servers. Confirm specific tools during scoping.
Usually not to start. We work with your identity provider, cloud IAM and a secrets vault you already have where possible. If a dedicated NHI or secrets tool would help, we'll say so with the cost, and you decide.
Those are platform features for giving agents and workloads proper identities and applying conditional access to them. We help you use them well — blueprints, owners, scopes and policies — and cover the identities that live outside Microsoft too.
No single service does. Our outputs provide evidence for access control, logging and AI system inventory that DPDP safeguards and ISO/IEC 42001 expect. XcellHost is not a certification body and does not give legal advice; we prepare you for certification by an accredited body.
We design the work so we don't need to. Scanners report the location and type of a secret with its value masked, and rotation is done by your team or through the vault with logged, time-limited access.
Typically 3–5 weeks for the standard scope of up to 3 agents and one cloud tenant*, including the re-test window. Larger estates are planned in phases.
In Indian rupees with a GST tax invoice: the assessment as a one-time fee, managed governance monthly or annually (annual gets 10% off the fee*). Prices on this page are indicative until quoted.