Security at Mobasi
Secure AI for sensitive
investigations.
Your evidence stays under your control. Mobasi is built for sensitive investigations with a zero-data-retention posture, transparent agent actions, and controls that stand up to enterprise review.
Investigative work demands controls that hold up to subpoena, audit, and adversarial scrutiny. Four principles shape how Mobasi is designed and operated.
Evidence stays under customer control
Mobasi runs against extractions, disk images, and case files where they already live. The full evidence container, the source of truth, never moves into a Mobasi-controlled store.
Zero data retention is the default
Prompts and model responses exist in memory only for the duration of the request. Nothing is persisted to a Mobasi database. Model providers do not retain content and do not train on it.
Every result is inspectable and reproducible
Agent actions are logged. Outputs are written as on-disk artifacts your team, and any reviewer, can open, re-run, and verify. No black-box conclusions.
Built for enterprise security review
Branch protection, gated release processes, vulnerability management, and vendor risk tracking are part of how Mobasi is operated and how it is delivered to customers.
Architected for security.
Mobasi runs from inside your environment and reaches out to the model and analysis services it needs. Nothing sits in the middle holding your evidence.
The runtime executes from the customer environment and reaches out to approved providers for inference and analysis. Outputs are written back as on-disk artifacts the investigator owns from the moment they exist.
Zero data retention.
Prompts and model responses pass through memory and are discarded when the request finishes. We don't write them to a database, log their contents, or send them to model training.
When an investigator runs a query in Mobasi, the prompt and the model response exist in memory only for processing. Nothing is persisted to a Mobasi-controlled database. No content is written to logs. Our model providers do not retain the data and do not use it to train or improve models.
Mobasi staff cannot access prompts or responses because the data is never stored anywhere reachable. Only you can see your investigations.
Even an adversary with full access to Mobasi's infrastructure would find no investigation content to take. There is nothing stored to exfiltrate.
Anthropic
OpenAI
Google Gemini
AWS
AzureEvidence files stay on your machine.
Extractions, disk images, mailboxes, and case files stay on the workstation or storage you already control. The agent reads them where they are instead of uploading them.
When the agent needs to reason about a specific record, like a chat thread, an email header, or a file path, only that fragment is transferred for in-memory processing. It is never persisted on our side. The full evidence container, the source of truth, stays local to your environment for the entire investigation.
Transparent, inspectable, and reproducible.
Every agent run leaves a trail you can read: what it looked at, what it ran, what it concluded. Outputs are plain files on disk that another investigator can open and re-run.
Every agent run produces an audit trail of the actions it took: what it read, what tools it invoked, what it concluded, and in what order. Reviewers do not have to take an answer on faith.
Outputs are written as on-disk artifacts: reports, transcripts, methodology notes, and the underlying commands. Another investigator can open them, re-run the work, and verify the result.
This is the deliberate inverse of black-box AI. Speed without defensibility is not useful for the matters Mobasi is built for.
FAQs for security teams
Does customer evidence leave our environment?
Mobasi is built so the underlying evidence stays where it already lives: investigator workstations, agency storage, or customer-controlled infrastructure. The agent reasons against evidence in place rather than ingesting it into a vendor data lake.
Do you retain investigation data?
Zero data retention is the default posture. Prompts and model responses are not written to a Mobasi database, are not logged as content, and are not used to train models.
Can we see exactly what the agent did?
Yes. Agent actions are recorded as an audit trail, and outputs are written as inspectable artifacts so a reviewer, opposing counsel, or another investigator can trace how a finding was produced.
How do you support defensibility?
Findings are designed to be reproducible. The commands, inputs, and methodology behind a result are captured so the work can be rerun and verified independently of Mobasi.
Do you use a centralized evidence database?
No. The current architecture does not centralize evidence. Customers connect to model and analysis services directly from their own environment.
How do you handle AI-specific risks like prompt injection?
Prompt injection, tool extraction, and credential exposure are explicit design considerations. We treat them as ongoing engineering work rather than solved problems, and we are transparent about that with security teams under NDA.
Are you SOC 2 certified?
A SOC 2 Type 1 audit is in progress, with evidence collection and vendor documentation underway. We share current status, scope, and timeline directly with customers under NDA. We do not claim a completed certification we have not yet earned.
Can customers run on their own compute?
We have active conversations with customers who want customer-controlled compute. Specifics depend on the deployment and the matter. Reach out to security@mobasi.ai and we will walk through what is supported today.
Send your security team.
We answer questionnaires, share documentation under NDA, and walk through the architecture in detail.