Skip to main content

Security & Trust

Security underpins AISafe. You need to trust that the platform protects your data, that the AI agents cannot cause harm, and that the platform itself stays secure when you use an AI-powered tool to audit your source code and test your applications.

The platform isolates agent execution in sandboxes and handles your source code and data throughout the assessment lifecycle. Email [email protected] for information not covered here.

Key principles​

  • Isolation by default: all AI agent execution happens in isolated sandboxes with controlled network egress. Agents cannot reach unintended services or exfiltrate data.
  • No persistent credential storage: source access tokens expire within a short window, scope to a single repository, and do not touch disk.
  • Tenant isolation: the platform scopes your data to your organization. Cross-organization access returns 404 to prevent information leakage.
  • Strict input validation: all API requests reject unknown fields, preventing injection of server-managed values.
  • No model training on your data: AISafe does not use your source code, findings, or assessment data to train AI models. The platform uses LLM providers for inference only.
  • Audit trail: the platform records key actions (assessment creation, finding triage, member changes, integration configuration) in an append-only audit log accessible to organization admins. Organizations with audit-chain integrity enabled can verify the retained trail against a tamper-evident per-organization hash chain; retention purges are sealed with signed checkpoints.

Compliance & certifications​

AISafe is building toward industry-standard compliance frameworks. The programs below are in progress. Contact us for the current status of any item before relying on it in a procurement review.

  • SOC 2 Type II: in progress. Reach out to request the latest status or a trust-center walkthrough.
  • GDPR: we support standard data processing terms and EU data residency. Contact us for a DPA.
  • HIPAA-ready: the architecture supports HIPAA-aligned workloads. Contact us to discuss a BAA and configuration for regulated data.

We update this page as each program reaches completion. Do not assume a certification is active unless we list it as completed here.

If you need a specific attestation, questionnaire response, or vendor risk assessment, email [email protected] and we will route it to the right team. We maintain an up-to-date security questionnaire package and can schedule a live review with our security team on request.

Data residency​

The platform processes your data in European Union regions. Source code you upload for a code audit stays encrypted at rest until the retention period configured for your organization expires, at which point the platform deletes it. Assessment artifacts (findings, reports, traces) follow the same retention and deletion policy. You can request early deletion of specific assessments at any time from the dashboard or via the API.

Encryption applies both in transit (TLS 1.2+ for all connections) and at rest. Under the principle of least privilege, the platform restricts data access to authorized personnel and logs all access for review.

The platform encrypts operational recovery backups and holds them within the same EU region. Backup retention follows the deletion policy described above.

If your organization has specific residency requirements beyond the default EU processing region, contact [email protected] to discuss available options before you begin uploading sensitive data.

For details on how source code and assessment data move through the platform, see Data handling.

Responsible AI​

AISafe uses large language models for inference only. We do not train, fine-tune, or improve models on your source code, findings, or assessment data. No model, ours or our providers', trains on your content.

The platform logs each action an AI agent takes during an assessment and makes it auditable. You can review the full step-by-step trace of each agent run, including tool calls, inputs, outputs, and decisions, from the assessment detail view or the research trace visualizer. Organization admins can export these traces for compliance review or incident investigation.

The sandbox environment constrains agent behavior: the platform controls network egress, scopes file system access, and validates tool invocations before execution. Agents cannot escalate privileges or access credentials belonging to other tenants.

If you need to verify the inference-only commitment as part of a vendor review, request a written attestation from [email protected].

For details on how the platform constrains agent execution, see Sandbox isolation.

Disclosure & vulnerability reporting​

This section is our coordinated vulnerability disclosure (CVD) policy. It tells you what you may test, what protection you have when you test in good faith, and what we commit to in return.

Report vulnerabilities to [email protected]. The same address is published in machine-readable form at https://aisafe.io/.well-known/security.txt, per RFC 9116.

Scope​

Everything we operate under the aisafe.io domain is in scope, including all subdomains.

The following are out of scope:

  • Denial of service and volumetric testing, including load, stress, and resource-exhaustion testing against any environment.
  • Social engineering of our staff, contractors, or customers, including phishing and pretexting, and any physical attack against our offices or personnel.
  • Third-party services we do not operate, such as our upstream vendors and LLM providers. Report those to the vendor directly. If the issue is in how AISafe integrates with a third party, that is in scope.
  • Findings with no demonstrated security impact, such as raw automated-scanner output, missing hardening headers, or weak TLS cipher suites with no exploit path. Show the impact and we will look at it.

Rules of engagement​

Test only against organizations and accounts you own. Do not access, modify, or delete data belonging to another tenant, and do not degrade the service for other users.

Stop as soon as you have proven the issue. Do not pivot further into the system, escalate beyond the minimum needed to demonstrate impact, or pull more data than a proof-of-concept requires. If you encounter another customer's source code, findings, or personal data, stop immediately, do not retain a copy, and tell us in your report.

Safe harbor​

If you make a good-faith effort to follow this policy, we consider your security research authorized. We will not initiate legal action against you, we will not refer you to law enforcement, and if a third party brings a claim against you for research conducted within this policy, we will make it known that your activity was authorized.

If you break something accidentally while acting in good faith, tell us. We will treat it as part of the research, not as an attack. We interpret this policy in your favor when a case is genuinely ambiguous, so a good-faith mistake does not cost you the protection above.

This safe harbor covers only the systems we operate. It cannot waive the rights of third parties, so it does not authorize testing against our vendors or against another customer's environment.

How to report​

Include the following where possible:

  • A description of the issue and its potential impact.
  • The affected product area (API, dashboard, agent execution, integrations).
  • Step-by-step reproduction instructions or a proof-of-concept.
  • Any relevant timestamps, request identifiers, or account identifiers.
  • The account or organization you tested from, so we can separate your testing from real activity in our logs.

What we commit to​

  • Acknowledgement within one business day of receiving your report.
  • A triage decision and a remediation timeline within five business days, once we have confirmed the issue. We prioritize critical issues and ship fixes for them in an expedited release cycle.
  • Progress updates at each significant step, and confirmation when the fix ships.
  • Coordinated publication within 90 days of your report. We will agree a disclosure date with you and aim to publish sooner when the fix lands earlier. If a vulnerability is being actively exploited, we may act faster, and we will tell you if we do.
  • Credit for the researchers who report valid issues, unless you would rather stay anonymous.

Please do not disclose a vulnerability publicly before that coordinated date. We do not operate a paid bounty program at this time.

For security questions about architecture, access controls, or integration security, email [email protected] and we will route your inquiry to the right team.

Topics​