Service account bypassed row restrictions
The AI query path used a privileged service account instead of enforcing the requesting user's row-level access.
Back to Use Cases
10Developer / Data PlatformsAthena + AchillesEnterprise AI Deployment Assurance
Data Platform / Cloud AI Integration
A data platform or cloud AI integration connects AI to the systems where business data, logs, analytics, models, metadata, and workflows live.
Buyer question
Can this AI connect to cloud and data platforms without bypassing permissions, exposing sensitive data, trusting poisoned metadata, or routing information through unapproved models?
Scenario
Scenario overview
A company wants AI to work with real enterprise data. The AI may connect to warehouses, lakehouses, cloud storage, dashboards, metadata catalogs, vector indexes, model gateways, logs, notebooks, APIs, data governance tools, or workflow systems. This can unlock powerful analytics and automation. It can also create new data paths, permission issues, model-provider exposure, and audit gaps.
Why it matters
Why this matters
When AI connects to data platforms, it can reach the crown jewels of the business. Authorization must happen before data reaches the model, not after. Metadata, logs, dashboards, and notebooks can also become prompt injection surfaces.
Risk surface
What can go wrong
- AI service accounts bypass user row-level restrictions.
- Generated SQL selects restricted columns.
- Vector indexes include restricted documents without ACL metadata.
- Table comments or metadata influence AI answers.
- Sensitive cloud logs enter model context.
- Fallback model routes lack approved handling.
- AI analytics summaries overstate causality.
- Query and retrieval audit logs are incomplete.
Assessment scope
What Mythos reviews
- Cloud and data architecture
- Warehouses and lakehouses
- Storage buckets
- Catalogs and dashboards
- Logs and notebooks
- Vector indexes
- IAM and service accounts
- Row-level and column-level security
- Model gateway routes
- SQL generation
- Retrieval permissions
- Exports and tool actions
- Prompt injection through metadata
- Audit logs
Mythos projects
Projects assigned
Athena
maps cloud systems, data reach, IAM, service accounts, sensitive data, model routes, logs, metadata exposure, and evidence.
Achilles
tests query behavior, retrieval boundaries, prompt injection resistance, output quality, model route behavior, export controls, and release readiness.
Minotaur
may support internal-only adversarial scenarios involving poisoned metadata, malicious table comments, unsafe SQL prompts, and cloud-tool misuse.
Illustrative findings
Example findings
Illustrative examples of what a Mythos assessment may surface. They are representative patterns, not findings from a specific customer.
Generated SQL selected restricted columns
The AI generated a query that included sensitive columns outside the user's allowed view.
Metadata prompt injection influenced answer
A table comment included instruction-like text that changed the AI's response.
Sensitive cloud logs entered model context
Logs containing secrets and internal architecture were retrieved into the model context.
Incomplete AI execution receipts
The system could not reconstruct which data, query, model route, and user context produced a given output.
Deliverables
What the customer receives
- Cloud and data architecture map
- Data reach and sensitivity matrix
- Identity and permission report
- Model route and provider boundary report
- Query and retrieval safety report
- Technical findings appendix
- Evidence pack
- Remediation backlog
- Retest plan
- Capability-by-capability release recommendation
Decision
Decision supported
Whether AI should remain sandboxed, operate over governed read-only data, generate SQL, use vector indexes, access sensitive datasets, export results, or use cloud/data tools after remediation and retest.
Recommendation
Final recommendation
A cloud-connected AI system should not expand until the organization can prove what data is reachable, which permissions apply, where model routes go, and what evidence supports each output. Mythos should help make enterprise data AI reviewable before deployment expands.
Mythos AI Security
Evidence-first AI deployment assurance.
Authorized. Scoped. Human-controlled.
Start the Assessment
Ready to review a system like this?
Tell Mythos what you are building, connecting, or preparing to release. We will help identify the right assessment path.