[AI Security Posture Management for Banks]
Know every AI system in your bank. Prove it's safe.
We find every AI model, chatbot and data connection your teams use, check each one for the 10 most common AI risks, and give you reports your auditors and regulators accept.
- Runs inside your bank, data never leaves
- Nothing to install on servers
- Covers OWASP Top 10 for AI & NIST AI RMF
Enterprise AI security score
AI assets discovered
312
+14 this week
Open critical
7
−3 vs last week
MTTR (critical)
4.2d
−61% QoQ
Controls passing
91.4%
NIST AI RMF
Exposure heatmap
Business unit × OWASP LLM Top 10 (2025) · hover a cell
Payments · Agent holds write scope on payments API
Active RAG & agent connectors
23 total- 09:41:02
Confluence: Retail KB
RAG source
- 09:38:55
SharePoint: Credit Policy
RAG source
- 09:12:40
pgvector: kyc-embeddings
Vector store
- 09:42:11
Azure OpenAI: gpt-4o (EU West)
Model endpoint
- 09:40:27
Payments Agent: tool:transfer
Agent plugin
AEG-F-0412 · Agent write scope exceeds approved risk appetite
01 / The problem
Your current security tools can't see AI risks
Code scanners were built for ordinary software. With AI, the risk sits in what the AI is told, what data it can read and what actions it can take. Those tools never look there.
What your code scanners check
- Source code & open-source dependencies
- HTTP endpoints & authentication flows
- Container images & IaC drift
Where AI risk actually hides
- Hidden instructions given to the AI, and attempts to trick it
- Which internal documents the AI can read, and who it shows them to
- What AI agents are allowed to do on their own, like move money
- Where the AI model came from and what data trained it
No one owns AI risk
Data scientists build the AI, security checks the code, risk keeps spreadsheets. Nobody sees the full picture, so it lands on the CISO's desk when the auditors arrive.
Sample data · illustrative
- ML Platform
- Jupyter, MLflow
- AppSec
- SAST / DAST
- Model Risk (SR 26-2)
- GenAI out of scope
- Named risk owner
- Unassigned
AI you don't know about
Teams connect chatbots to Teams and email, send data to public AI services and launch tools nobody registered. You can't protect what you can't see.
Sample data · illustrative
- api.openai.com egress
- 38 hosts
- Unregistered vector DBs
- 6 clusters
- Browser AI extensions
- 212 seats
- Registered in CMDB
- 41%
Risky third-party parts
Every AI app is built from outside models, libraries and plugins. One plugin with too many rights can turn a simple chatbot into a door to your core banking system.
Sample data · illustrative
- langchain 0.0.131
- CVE-2023-29374
- Model weights
- Unsigned
- Third-party plugins
- 17 in use
- Vendor DPAs on file
- 9 / 14
02 / Platform
One place to find, check and fix AI risk
Not just another chatbot filter. A live map of all AI in the bank, checked around the clock against the rules your regulators expect.
AI asset inventory
Discovered via AWS, Azure, M365, K8s| Asset | Type | Owner | Permissions | Status |
|---|---|---|---|---|
| retail-assist-prod | LLM app | Digital Channels | read:kb | Sanctioned |
| gpt-4o · eu-west | Model endpoint | AI Platform | none | Sanctioned |
| kyc-embeddings | Vector store | Onboarding | read:pii | Review |
| payments-ops-agent | Agent | Payments Ops | write:transfer | Violation |
| credit-memo-summarizer | LLM app | Unassigned | read:sharepoint | Shadow AI |
| llama-3.1-70b-ft-aml | Fine-tuned model | Financial Crime | none | Validating |
Models
47
RAG sources
88
Agent tools
61
03 / OWASP mapping
Checked against the 10 most common AI risks
OWASP Top 10 for AI is the industry list of the biggest AI security risks. We test for each one and link every result to a control your auditors can check.
- OWASP categories
- 10 / 10
- Automated checks
- 240+
- NIST AI RMF functions
- 4
- Initial assessment
- < 24h
- LLM01Prompt InjectionAdversarial prompt suites, indirect-injection canaries in RAG sourcesAI-IV-01
- LLM02Sensitive Information DisclosureOutput DLP for PAN, IBAN, SWIFT/BIC and national IDsAI-DP-03
- LLM03Supply ChainModel SBOM, weight signing, framework CVE correlationAI-SC-02
- LLM04Data and Model PoisoningTraining & fine-tune data provenance attestationsAI-DI-04
- LLM05Improper Output HandlingTaint tracking from model output to sinks (HTML, SQL, shell)AI-OH-01
- LLM06Excessive AgencyAgent tool-scope analysis vs. approved risk appetiteAI-AC-07
- LLM07System Prompt LeakageSystem-prompt extraction probes, secret scanning in promptsAI-CF-02
- LLM08Vector and Embedding WeaknessesVector-store ACL and cross-tenant retrieval testingAI-DP-05
- LLM09MisinformationGroundedness scoring against approved source corporaAI-QA-01
- LLM10Unbounded ConsumptionRate-limit, token-quota and cost-anomaly policiesAI-RS-02
04 / Solutions for banks
Reports your auditors and regulators accept
Always mapped to the main AI rulebooks for banks (NIST AI RMF, FS AI RMF, EU AI Act). Get a complete audit pack in one click and connect the tools you already use.
Framework coverage
Sample dataFS AI RMF
88%
Financial services AI risk management
NIST AI RMF 1.0
91%
Govern · Map · Measure · Manage
NIST AI 600-1
84%
Generative AI profile
ISO/IEC 42001
79%
AI management system
EU AI Act
72%
High-risk system obligations
DORA
81%
ICT third-party & resilience
NIST · GOVERN
94%
NIST · MAP
92%
NIST · MEASURE
87%
NIST · MANAGE
90%
Evidence package
Sample dataAUD-PKG-2026-Q4-017
Prepared for Internal Audit · ECB JST review
- control-test-results.csv1,284 rows
- owasp-llm-mapping.pdf42 pp
- risk-register-snapshot.json148 risks
- model-inventory-sbom.cdx.json312 assets
- approver-signoffs.pdf6 signatures
SHA-256 · WORM-sealed
9f2c4e81a07d3b5e6c18f4a92d0b7e3c5a61f8d24b9e0c7a3d5f1e8b6c2a4d90
Works with the tools you already have
Bi-directional · API & webhookGRC
ServiceNow IRM · Archer · MetricStream
SAST / SCA
Checkmarx · Semgrep · Snyk
DAST
Burp Suite Enterprise · Invicti
CMDB & ITSM
ServiceNow CMDB · Jira
SIEM / SOAR
Splunk · Microsoft Sentinel
Identity
Entra ID · Okta · CyberArk
Prompts never leave your perimeter
Collectors analyse content in-VPC and export only metadata and verdicts.
Customer-managed keys
BYOK via AWS KMS, Azure Key Vault or on-prem HSM (FIPS 140-2 L3).
Read-only by design
Least-privilege, scoped service principals. No write access to production.
05 / Savings
See how much it saves you
Move the sliders to match your bank. The numbers update instantly, and every assumption is shown.
Estimated value per year
$832K
$2.50M over a 3-year term
- Breach losses avoided$306K
- Less manual review work$241K
- Faster fixes$284K
Days to fix
28d → 11d
Staff hours saved / yr
3,624
›Model assumptions
- Incident likelihood reduction: 42%
- Review effort: 64h / app / yr, 65% automated
- Findings: 14 / app / yr, 3.5h saved each
- MTTR reduction: 62%
06 / Security & deployment
Runs inside your bank. Your data stays there.
Pick how you want to run it. Every option has the same features.
Dedicated SaaS
Single-tenant, in-region (EU, UK, US, APAC). Isolated compute and storage per customer.
Private cloud
Deployed into your AWS, Azure or GCP account via Terraform. Your VPC, your keys, your logs.
On-premise / air-gapped
Kubernetes or OpenShift with offline update bundles. No outbound connectivity required.
AI Security Posture Management (AI-SPM) is the continuous process of finding every AI system an organization uses, checking each one for security and compliance risks, and tracking those risks until they are fixed. kulrus does this for banks: it builds an inventory of AI models, chatbots, data sources and agents, tests them against the OWASP Top 10 for LLM applications, and produces audit-ready reports for NIST AI RMF and FS AI RMF.
A prompt scanner checks individual messages sent to one AI model. kulrus looks at the whole picture: which AI systems exist across the bank, what data and tools each one can reach, who owns each risk and whether it has been fixed. It is built for CISOs, AI governance and risk teams who need to prove control to auditors and regulators.
No. Collectors inspect content inside your environment and export only metadata, verdicts and hashes. In private cloud and on-prem deployments, nothing leaves your network at all, including telemetry, which can be fully disabled.
Role-based access control with SSO (SAML 2.0 / OIDC) and SCIM provisioning. Built-in roles for AppSec, ML Engineering, Model Risk, GRC and Internal Audit, scoped by business unit, legal entity and environment. Sensitive fields can be masked per role.
Every action (finding triage, risk acceptance, control change, export) is written to an append-only, hash-chained log. Logs can be streamed to your SIEM and archived to WORM storage to satisfy record-keeping requirements.
Azure OpenAI, AWS Bedrock, Google Vertex AI, Anthropic and OpenAI APIs, self-hosted open-weight models (vLLM, TGI, Ollama), and orchestration via LangChain, LlamaIndex, Semantic Kernel and custom agents. Vector stores include pgvector, Pinecone, Weaviate, Elastic and OpenSearch.
No. Discovery is agentless, using read-only cloud, identity and SaaS APIs plus optional network egress logs. An optional SDK or gateway integration adds runtime telemetry where you choose to enable it.
Most banks complete initial inventory and OWASP assessment within the first week. Evidence packages for NIST AI RMF and FS AI RMF are generated on demand, signed and versioned against a point-in-time snapshot.
See every AI system in your bank in under a week.
A 45-minute call with our banking team. We show you what we find, check one of your AI projects against the top 10 AI risks, and hand you a draft report for your next audit.