Security

The authority to act is enforced outside the model.

Your data lives in infrastructure you run, your secrets stay out of the model’s prompt, and an action above your approval threshold waits for you. The model you choose still sees the prompts you send it.

Every control below says where it stops. The gaps are tracked in public, in a residual-risk roadmap.

01Shared responsibility

Sophon ships the controls. The deployment is yours.

Our own assessment of who owns what. Where something is Enterprise-only or has an exception, the row says so.
Where it runs
Sophon shipsSelf-hosted with Docker Compose or Kubernetes and Helm. Conversations, memory, documents, and credentials are stored only in your deployment.
You ownThe region, cloud, or on-premises site you choose, which is how data-residency requirements are met. Host-OS hardening and physical security.
Models
Sophon shipsModel-agnostic: Anthropic, OpenAI, Gemini, Mistral, xAI and 20+ other provider types, or local models via Ollama, LM Studio or any OpenAI-compatible server; none configured by default. Prompts go only to the providers you configure.
You ownWhich provider sees your prompts, and that provider's own security. The model you choose still sees the prompts you send it.
Outbound connections
Sophon shipsNo usage analytics and no behavioral tracking. License validation is offline. The marketplace client checks for updates, reports anonymous install counts, and downloads a package only if you turn automatic updates on.
You ownWhether the marketplace stays enabled. Disabling it stops all three.
Sophon shipsAn encrypted vault. The runtime resolves a credential at call time, so it never enters the model's prompt.
You ownThe scopes you grant at connect time, and which skills you install and connect. A skill that declares a credential receives it.
Sophon shipsNormal approval modes ask at or above your threshold and always at Critical. A separate host-enabled, explicitly confirmed Allow all mode can cover Critical calls. A timeout or cancel is never consent.
You ownThe threshold, per-tool overrides and tenant membership. Workflow cards are owner/Admin-only; chat-requested agent installs need the full Dashboard review.
Sophon shipsSkills run in Docker with CPU, memory, and time limits and no default egress. Marketplace skills pass static analysis and human review, and every install is checksum-verified.
You ownConfiguring gVisor, keeping process fallback off outside development, and every plugin you enable: plugins run full-trust, outside the sandbox. Third-party packages beyond first-party review.
Sophon shipsEnterprise: SSO via OIDC (Entra ID, Okta, Google Workspace, Auth0, or any OIDC provider) or SAML, and RBAC with custom roles for who can create agents, install skills, approve actions, access memory, or manage channels.
You ownMFA comes from your identity provider; admins can require TOTP two-factor for local accounts. Session policies, conditional access, role assignments and access reviews.
Sophon shipsEnterprise: each tenant's users, agents, memory, documents, connections, and audit trail are separated at the data layer by query filters. Vault entries are tenant-scoped, and a per-tenant vault backend is available.
You ownWhether tenants share a vault backend or each get their own.
Sophon shipsThe technical controls SOC 2, ISO 27001, HIPAA, and GDPR require, plus compliance reports and audit export on Enterprise.
You ownCertification. It applies to your deployment; Buildersoft holds no attestation of its own.

02Credentials

The model never sees the secret.

Step through one call to a sandboxed skill and watch where the credential is. The model's prompt holds the request and the result, never the key. Switch to the roadmap to see the broker that is planned but not shipped.

Credential path, one sandboxed call

  1. Model

    the prompt

    never holds it
  2. Vault

    runtime, encrypted at rest

    credential (holds the credential at this step)
  3. Sandbox

    the skill's container

  4. API

    the external service

Model prompt

  • ask: create the calendar event
  • credential: not here

The model proposes a tool call. Its prompt holds the request and, later, the result. The credential is part of neither.

1 / 5

Built-in tools skip the sandbox: the runtime makes the call outside the model context and returns only the sanitized result.

Where this stops: a compromised skill can read the credential it was given, and what it receives is the stored credential, not a short-lived scoped token. A broker that keeps the secret outside the sandbox is on the roadmap. Grant a skill only the connections it needs.

AuthorityRoom 04 of 13

03Approvals & risk

You decide what runs without asking.

Every tool call carries a risk level. You set the threshold, from asking on every call to asking only at Critical. It ships at Medium; a timeout or a cancel is never recorded as an approval, and Sophon never assumes consent.

Since 2.2, a separate Allow all tools mode can cover Critical calls too. It is off unless the host enables it; an Admin in a trusted workspace must confirm it for a set duration, and reading untrusted content ends it. This is separate from ordinary risk thresholds. Read the permission-mode limits. Marketplace agent installs requested in chat always require the full Dashboard review.

Medium (default)

Critical always asks, whatever the threshold says.

Runs3

  • memory.searchNone
  • web.searchLow
  • document.extractLow

Asks you5

  • calendar.create_eventMedium
  • message.sendMedium
  • gmail.sendHigh
  • workflow.createHigh
  • system.executeCritical

Threshold Medium. 3 of 8 calls run, 5 ask you.

Recreation of Settings › Security › Approvals. Risk levels as declared by each tool.

NoneSearch memory, check the timeRead-only, cannot affect anythingRuns
LowWeb search, open a page, extract a documentReads with side effects, minor writesRuns (below the default threshold)
MediumSend a channel message, write a memory, create a calendar eventWrites data, sends messages you could undoAsks you (the default threshold sits here)
HighSend an email, create a workflow, run code in the sandboxExternal, costly, or hard to reverseAsks you unless you raise the threshold
CriticalRun a shell command on the host, forget all memoryDestructive or privilegedAlways asks (no threshold or override lowers it)

Where the gate stops

  1. Approval is a ceiling you set, not a switch the model flips. At the shipped Medium threshold, None and Low calls run without asking.
  2. Unattended runs, including heartbeat, cron, subagent, and webhook, proceed past the plan gate by design. Each step still asks at your threshold, and Critical always asks.
  3. A heartbeat with auto-approve on approves Medium and below for itself, so it can send messages, including email, on your behalf. High and Critical always page you.
  4. In a multi-tenant deployment, any user in a tenant can answer any pending request.
  5. There is no bulk-operation threshold: many individually low-risk calls are not aggregated or gated as a batch (R-12).

How approvals work

04Defense in depth

Seven layers, each with its edge marked.

Each layer is documented, including where it stops.
  1. Encrypted at rest: DPAPI-bound to your account on Windows, AES-256 under a local key file on macOS and Linux. HashiCorp Vault, AWS Secrets Manager, and Azure Key Vault on Enterprise.

    Where it stops

    On macOS and Linux the local key is a file on disk, not the OS keyring (R-06). It does not keep a credential away from a sandboxed skill that declared it (R-13).

  2. Integrations prefer modern OAuth with proof keys. The token carries only the scopes you granted at connect time, and the runtime renews expired access tokens from the stored refresh token. Static credentials are a last resort.

    Where it stops

    Scope is fixed when you connect, not per call. Sophon does not mint short-lived or per-call scoped tokens; what is delivered is what is stored.

  3. Every tool call is scored from None to Critical. You configure which levels run and which ask, down to the individual tool. The threshold ships at Medium, and Critical asks whatever you set it to.

    Where it stops

    No bulk-operation threshold yet: individually low-risk calls are not gated as a batch (R-12).

  4. Every skill, bundled, marketplace, or self-authored, runs in a Docker container with CPU, memory, and time limits. gVisor kernel isolation applies where the runtime is configured.

    Where it stops

    Plugins are full-trust local processes, not sandboxed code. If Docker is unreachable and you enable process fallback, code runs as a child process with host network access.

  5. Sandboxed skills have no network access by default. A skill reaches the network only through an allowlist declared in its manifest, and network tools are approval-gated.

    Where it stops

    Python that imports third-party packages enables network for that run to install them. Exfiltration to an allowlisted or public endpoint is not detected; there is no egress DLP (R-07).

  6. Content that did not come from you is wrapped, scanned, and recorded. Off, Warn, or Block, admin-set for the host and shipping as Warn. Block stops a turn only on third-party content: a tool result, a webhook payload, a message from someone else.

    Where it stops

    It matches known phrasing, so treat it as a tripwire, not a boundary: obfuscated, split, or image-embedded instructions get through, and tools called through the embedded MCP server are not scanned (R-04).

  7. Every action is logged: who triggered it, what happened, when, and the outcome. Enterprise adds compliance export, retention policies, and tamper-evident storage.

    Where it stops

    Tool parameters are hashed, so content-level exfiltration is not visible (R-02). Auth-failure and webhook signature-mismatch logging is partial (R-10).

05Threat model

Mapped to MITRE ATLAS, gaps included.

Sophon's controls are mapped to MITRE ATLAS, the industry framework for adversarial threats against AI systems. A control is listed only where it is implemented today; partial ones point to the residual-risk roadmap.

15 ATLAS tactics covered34 techniques mapped10 trust boundaries

ExecutionIndirect prompt injectionAML.T0051Content wrapping + Warn/Block guardResidual: R-04, R-07
ExecutionMalicious skill codeAML.T0011Docker + gVisor sandboxResidual: R-11
Credential AccessCredential theftAML.T0055Encrypted vault, out of the promptResidual: R-06, R-13
ImpactIrreversible actionAML.T0048Human approval gatesResidual: None listed

06Supply chain

Check what you install. Our installers are not code-signed yet.

Installers we host list a SHA-256 checksum on their download page, so you can verify what you downloaded before you run it. Code signing is on the roadmap, not in the product.
Installers
SHA-256, not code-signedInstallers we host list a SHA-256 checksum on their download page. The installers are not code-signed yet, so Windows SmartScreen and macOS Gatekeeper will warn the first time you run one. Signing is on the roadmap; until then the checksum is the check that the file is the one we built.
Sophon Node
SHA-256, not code-signedNode binaries are not code-signed either, so Windows and macOS treat them as software from an unidentified developer. Check the checksum before you run one.
Marketplace skills
Reviewed, checksum-verified, not signedAutomated static analysis and a mandatory human review of every version before publication. Every install is verified against a SHA-256 checksum and fails closed on a mismatch. Packages are not cryptographically signed yet (R-11).
Plugins
Full-trust, not signedNo container, no code signing, no resource limits. A plugin can do anything the Gateway's OS user can do, so install them from sources you trust.
Third-party packages
Out of scopeThe threat model covers first-party review. Supply chains of third-party packages beyond it are out of scope.

07Vulnerability disclosure

Found something? Tell us privately first.

Report a vulnerability or a new threat that affects Sophon before disclosing it publicly. Reproducible reports with a clear impact statement are appreciated and acknowledged.

Write to

security@buildersoft.io

Include

  • A description of the vulnerability
  • Steps to reproduce it
  • Your assessment of the impact
  • How to reach you
Acknowledgement
Within 48 hours.
Coordinated disclosure
Typically 90 days.
Keep it private
Not in public GitHub issues and not through the support form. Share proof-of-concept material privately so a fix can ship before details circulate.

The full policy is in the threat model. For anything that is not a vulnerability, use support.

08Documents

What your security review can read today.

Public documents are linked. Anything else is on request through sales, and we will tell you plainly whether it exists.
  • MITRE ATLAS threat modelTrust boundaries, technique coverage, detection coverage, and the residual-risk roadmap.PublicOpen MITRE ATLAS threat model
  • Security documentationCredential vault, sandbox, prompt-injection defense, rate limiting, approvals.PublicOpen Security documentation
  • Privacy policyWhat Buildersoft does and does not collect.PublicOpen Privacy policy
  • Terms of serviceThe terms that govern your use of Sophon software and related services.PublicOpen Terms of service
  • End user license agreementThe software license under which Sophon is used.PublicOpen End user license agreement
  • Compliance and audit reportsUser access review, permission matrix, approval history, credential access, and more, generated by your own Enterprise deployment for your auditors.Your deploymentOpen Compliance and audit reports
  • Security review materialAnything else your review needs. Ask, and we will tell you plainly whether it exists.On requestRequest Security review material
  • Third-party attestationBuildersoft holds no SOC 2 report, ISO certificate, or other attestation of its own.Not held

Sophon ships the technical controls SOC 2, ISO 27001, HIPAA, and GDPR require. Certification applies to your deployment; Buildersoft holds no attestation of its own.

Bring us your security review.

Tell us how many people, on whose infrastructure, and what your reviewers need to see. We will walk through the controls, the limits, and the Enterprise tier with you.