Disclosure
Report security issues to security@mindburn.org. Do not include secrets or private customer data in public issues.
Security
This page explains website security, responsible disclosure, and company-level boundaries. HELM technical details live in HELM docs.
Posture
Report security issues to security@mindburn.org. Do not include secrets or private customer data in public issues.
Public forms and email are for triage only. Private evidence belongs in a reviewed channel.
CI checks npm audit, registry signatures, pinned actions, CodeQL, and SBOM output for the public site.
The public site does not use cookies or ad trackers. Optional web vitals are aggregate page-load metrics only.
Company boundaries
Company artifacts are permission-scoped.
Local knowledge state cannot authorize side effects.
Canonical knowledge promotion requires provenance.
Sensitive payloads stay off graph.
Query answers declare draft, canonical, disputed, or stale status.
Connector drift fails closed.
Unknown MCP tools quarantine before dispatch.
Code comments and README text are untrusted evidence, not instruction authority.
Adversarial classes
External tool output, MCP servers, webpages, documents, code comments, and README text are treated as untrusted evidence unless normalized and approved.
Unknown or unapproved MCP servers deny until approval receipt exists.
Hostile tool responses quarantine before they enter execution context.
Comments, README text, and forged approval text are evidence to scan, not instructions to follow.
Replayed intent, stale policy hash, invalid signatures, and schema mismatch fail closed.
Tenant and permission labels block data from crossing into another workspace.
Rate limits, high-risk quotas, deliberate approval UI, and summary-hash binding reduce reflexive approval.
Secret labels and hostile-output labels deny unsafe export paths.
Scope, idempotency, drift behavior, rollback, and receipt contracts keep compromised connectors visible.
Action boundary
HELM's public story is intentionally narrow: model proposals are not trusted as execution authority.
Category boundaries
Adjacent tools can be useful sensors, tests, filters, or downstream controls. HELM's claim is narrower: decide whether a proposed side effect may execute, then leave source-owned evidence.
Guardrails may inspect prompts, traces, model output, or tool-call text. HELM decides whether a proposed side effect has authority to run, returns ALLOW, DENY, or ESCALATE, and records the decision.
Orchestrators decide what an agent tries next. HELM sits at the action boundary so only authorized side effects dispatch.
Logs show what happened. HELM blocks or escalates before execution and emits receipt evidence for later review.
DLP can explain sensitive data movement. HELM controls whether an AI-driven action is authorized to move, expose, or write data.
Model filters reduce harmful model behavior. HELM governs tools, connectors, infrastructure, and business side effects after the model proposes an action.
Objection handling
Public competitor language is useful only when it clarifies the boundary. It does not become a HELM capability claim.
Agent runtime security and MCP guardrails
Ask whether the block is deterministic authority or semantic detection, and whether the customer can verify the verdict offline.
HELM positions the MCP/tool-call path as default-deny execution authority with signed receipts and EvidencePack refs.
Cloud-control enforcement
Cloud controls are downstream architecture. They do not sit between every AI proposal and side effect.
HELM decides whether the AI-driven action should happen, then downstream controls can enforce durable cloud posture.
DLP and insider data risk
Endpoint data movement is valuable context, not generalized action authority.
HELM can consume data-risk signals as policy context while receipts prove whether the action was allowed, denied, or escalated.
Autonomous offensive testing
Exploit-chain evidence shows what can go wrong. It does not itself govern runtime authority.
HELM turns findings into policy regressions and blocks unauthorized side effects at execution time.
Model safety, red teaming, and runtime model filtering
Model behavior risk is not the same as authority over tools, connectors, and infrastructure side effects.
HELM uses model-safety findings as evidence input while keeping action approval, verdicts, and receipts at the execution boundary.
Responsible disclosure
Email security@mindburn.org with a short description, affected URL, steps to reproduce, and impact. Do not include secrets, credentials, customer data, private documents, or regulated data.
Open security.txtArtifacts
The public SBOM and provenance manifest describe the mindburn.org packages and build output. They are not third-party assurance reports. They do not replace HELM technical docs.