Eve library

Bellona Remote Creative Control Plane — Privacy & Legal Review (Review-Prep, 180.C.27.10)

Guide

Revision 1

On this page

Status: 180.C.27.10 — privacy/legal review-preparation artifact.

Implementation date: 2026-06-21.

What this document is — and is not#

This is a review-preparation artifact: a structured privacy and legal analysis of the Bellona Remote Creative Control Plane, written by engineering to bring counsel and the data-protection function to a productive sign-off conversation. It documents what the system captures, the jurisdictions implicated, the privacy risks (notably the employee-monitoring misuse risk), and the already-implemented technical controls that mitigate each — citing the exact module that enforces each control.

It is not legal advice and is not a sign-off. Legal sign-off by qualified counsel (and, where applicable, a Data Protection Impact Assessment under GDPR Art. 35, a CPRA risk assessment, and works-council / employee-representative consultation in jurisdictions that require it) is an external step this document prepares for. Open legal questions are flagged inline as [Counsel decision]. Nothing here may be relied on as a determination that a given deployment is lawful; that determination is counsel's, informed by the concrete deployment's jurisdictions, data subjects, and lawful basis.

The Remote Creative Control Plane lets a remote operator (or an Oshun agent acting under that operator) drive creative tools — Blender, Unreal, a browser, the desktop — on a physically separate host machine, streaming the host's screen back and sending commands forward. That capability is powerful and inherently privacy-sensitive, which is why it ships with the controls catalogued below.


1. Remote screen capture#

What is captured#

A live (optionally recorded) video stream of the host operator's screen, plus OCR'd text from screenshots and captured artifacts. The screen of a creative workstation routinely shows far more than the work in hand: open chat windows, email, a second person on a video call, a document left visible, a browser with the operator's logged-in sessions.

Data subjects and jurisdictions#

  • The host operator — the primary subject; their screen, and by extension their activity, is captured.
  • Incidental subjects — anyone whose image, voice (via a dictation overlay), or personal document is on screen but who is not a party to the session.

Screen and (especially) audio capture is regulated differently across jurisdictions: EU/UK GDPR treat the captured imagery as personal data requiring a lawful basis and data-minimization; many US states (and all-party-consent states for any audio) impose consent requirements; recording another identifiable person without their consent raises image-rights and, for audio, wiretap-statute exposure. [Counsel decision] which lawful basis (consent vs. legitimate interest) and which consent model (operator-only vs. all-incidental-subjects) applies per deployment region.

Controls implemented#

  • Operator consent + visibility. A session that opens a real personal profile or a privileged surface is a consented, visible-streamed session (apps/bellona/remote-host/src/browser-adapter.ts privileged-profile session; the operator sees the live stream of what is shared).
  • Incidental-subject fail-closed redaction. Detected video-call windows, dictation overlays, bystander-camera regions, and visible documents are redacted unless explicitly consented, via libs/bellona/remote-protocol/src/incidental-consent.ts (27.14) resolveIncidentalSubjectDisposition — the default disposition is redact-no-consent, and a visible document is redacted even with a generic grant unless an operator explicitly opts that exact region in. The redaction is a real stream-privacy region the host masks before the frame is encoded.
  • Credential-entry blackout. While a secret is typed locally, the stream and recording are paused (credential-entry-mode.ts, 27.04), so a password is never in the captured video.

2. Browser session handling#

What is captured#

When a session drives a real browser, it operates inside the operator's own authenticated browser profile: the operator's logged-in cookies, saved sessions, and page content are present in the tab. The stream captures the rendered page; artifacts may capture scraped DOM text and traces. This is the most sensitive surface, because the operator's authenticated identity (their email, banking, internal tools) can be reachable from the automated tab.

Credential handling#

  • No automated credential entry on a real profile. The credential-entry privacy control (browser-adapter.ts assertCredentialEntryAllowed, 19.06) refuses automated typing into a password field on a real personal profile unless the operator has explicitly allowed it. An injected page instructing the agent to "auto-fill the saved password and submit" is refused at this gate.
  • Local secret entry is blacked out. When the operator must type a credential, credential-entry-mode.ts (27.04) pauses capture and drops keystrokes.

What page content cannot do#

Page content is treated as data, never control. The 19.08 prompt-injection corpus (browser-prompt-injection.ts) pins each malicious page to the host gate it must not defeat; the command-envelope schema (command.ts) makes "privileged ⇒ approval required" structural, so a page cannot flip a command's approval bit. Navigation/redirect/upload to a foreign origin is refused by the origin gates; cookie exfiltration to an attacker origin is refused by the navigation-origin policy.

Captured-content redaction#

Scraped browser content, traces, and logs pass through the unified redaction engine (redaction-engine.ts, 27.05): API keys, tokens, JWTs, bearer headers, and form-field secret values (e.g. a value="…" on a password input) are removed before the text reaches an artifact or the audit. [Counsel decision] whether, for some deployments, real-browser sessions must run only in an isolated throwaway profile (see browser-profile-isolation.md) rather than the operator's personal profile.


3. Employee-monitoring risk (the central misuse concern)#

The risk#

A tool that streams an employee's screen, records it, and logs every action could be repurposed as covert surveillance — an employer silently watching a worker. This is the highest-severity privacy risk in the system. In the EU/UK, covert monitoring of employees is tightly constrained (proportionality, transparency, and in many member states works-council consultation); the UK ICO's employment-practices guidance and EDPB Opinion 2/2017 on data processing at work both stress transparency and necessity. Several US states require notice of electronic monitoring (e.g. NY, CT, DE). Covert, surreptitious monitoring is the misuse this tool must actively resist, not merely discourage.

Explicit guidance#

Bellona is not a covert employee-monitoring tool and must never be deployed as one. It is built for a creative operator to delegate work on their own machine, with the operator a visible, consenting participant. Deploying it to watch an employee without their knowledge and consent is a misuse the organization is responsible for preventing, and may be unlawful. This guidance is normative for any deployment.

Controls that mitigate the misuse#

  • Review-only template. organization-template.ts (27.06/27.11) ships a review-only org template under which no mutating command may run and recording is forbidden — a posture for observe-only use that minimizes captured data. checkReviewOnlyCommand enforces it (any mutating command is refused).
  • Operator presence + consent are structural. Privileged/personal-profile sessions are visible-streamed and consented (§1, §2); they are not designed to run invisibly behind the operator's back.
  • Recording is deny-by-default and gated. Stream recording requires explicit organization permission and a recorder allowlist (organization-control.ts stream-access); it is not on by default.
  • Incidental subjects fail closed (§1) so even a recorded session does not silently capture third parties.
  • Audit transparency. Every command and approval is recorded in an append-only, hash-chained audit trail with full actor attribution (actor-attribution.ts, 27.13: operator → orchestrator → specialist → adapter), so monitoring activity is itself auditable and attributable — covert use leaves a trail.
  • Redaction-by-default of secrets/PII (§6 of the security review) limits what any captured text can expose.

[Counsel decision] the required employee-facing notice/consent text and whether works-council / employee-representative consultation is needed per deployment jurisdiction before any session that captures an employee's screen.


The operator initiates and sees the session; privileged and real-personal-profile surfaces are consented, visible-streamed sessions (browser-adapter.ts). Pairing a device and starting a session are themselves gated organization surfaces (organization-control.ts device-pairing, session-start).

incidental-consent.ts (27.14) implements a prompt-based consent flow for subjects swept into the capture who are not the operator:

  • On detection of a video-call window / dictation overlay / bystander / visible document, a consent request is raised (createIncidentalConsentRequest) tied to the exact region and subject kind, shown to the operator.
  • Until answered, the disposition is fail-closed: the region is redacted.
  • A grant is recorded as a durable who/when/scope record (RemoteIncidentalConsentRecord): who decided, when, and the scope (this-region vs this-session). A grant for one region/kind/stream/session can never authorize another (the disposition function refuses cross-region, cross-kind, cross-stream, and cross-session reuse), and grants expire.

[Counsel decision] whether operator-mediated consent on behalf of an incidental subject is sufficient in a given jurisdiction, or whether the subject's own consent is required (in which case the correct posture is to keep the region redacted).


5. Recording retention#

retention.ts (27.16) implements jurisdiction-aware retention defaults with a bounded operator override and automated expiry enforcement, across three data classes (audit, recording, artifact):

  • EU — GDPR storage-limitation (Art. 5(1)(e)): recordings/artifacts purged quickly; audit kept only as long as the integrity obligation requires.
  • UK — UK GDPR / DPA 2018 parity with the EU posture.
  • California — CCPA/CPRA bounded, disclosed retention; a middle ground.
  • US-federal — longer holds for classes that may be subject to litigation hold / regulatory audit.
  • default — a conservative organization baseline when no jurisdiction is asserted.

resolveRetentionPolicy produces the effective policy for a (jurisdiction, data class, override) tuple, and computeExpiry drives automated purge (now >= expiresAt). An operator override may shorten recording/artifact retention (data-minimization is always allowed) and may lengthen up to a jurisdiction maximum, but can never push audit retention below the integrity floor — the hash-chained audit must outlive the operational data it describes so the chain stays verifiable. [Counsel decision] the concrete day counts per jurisdiction and data class (the table ships documented defaults; counsel confirms them against the deployment's obligations).


6. Audit export (who may export, and redaction)#

  • Who may export. organization-control.ts (27.06) audit-export gates exports to an allowlist of participants; the restricted org templates (enterprise/education/review-only) pin audit-export to the administrator participant id (organization-template.ts), so an ordinary operator cannot export the trail.
  • Redaction on export. When the org's redactionRequired is set (the default), an unredacted export is refused — the export must be redacted. The audit content itself is scrubbed through the unified redaction engine (redaction-engine.ts, 27.05), and the credential-entry / incidental-subject paths never write the underlying secret or raw subject into the audit in the first place (payloadCaptured: false).
  • Who did what. Every audit entry carries the full actor-attribution chain (actor-attribution.ts, 27.13): the human operator at the root through the agent/adapter that executed, rendered as a readable delegation label — so an exported trail attributes each action without inventing attribution.

[Counsel decision] any data-subject access-request (DSAR) handling for an incidental subject who appears in a recording/audit, beyond the redaction and the erasure path (erasure.ts).


7. Cross-border artifact storage (data residency)#

organization-governance.ts (27.08) implements a data-residency policy that refuses storing artifacts/audit/recordings in a region the organization does not permit:

  • DATA_RESIDENCY_REGIONS enumerates concrete storage regions (eu-west, eu-central, uk-south, us-east, us-west, ca-central, ap-southeast), and DATA_RESIDENCY_REGION_JURISDICTION maps each to the retention jurisdiction that governs data stored there — so residency and retention stay consistent (an EU-only org's data is governed by the EU retention posture).
  • checkDataResidency / assertDataResidency refuse a storage target whose region is not in the org's allowedRegions. An EU-only organization (allowedRegions = [eu-west, eu-central]) handed a us-east bucket is refused (DataResidencyViolationError) under the default enforce: true posture.

This is the technical control that keeps a remote-control session's captured data from silently crossing a border the organization (or the law applicable to its data subjects) does not allow. Cross-border transfer of personal data out of the EU/UK additionally requires a transfer mechanism (adequacy decision, SCCs, or equivalent); the residency control keeps data in-region by default, and any deliberate out-of-region storage is a [Counsel decision] requiring the appropriate transfer basis.


Implementing modules (index)#

Topic Module TODO
Incidental-subject consent + fail-closed redaction libs/bellona/remote-protocol/src/incidental-consent.ts 27.14
Credential-entry blackout (stream pause, no keystrokes) apps/bellona/remote-host/src/credential-entry-mode.ts 27.04
Browser credential-entry privacy control apps/bellona/remote-host/src/browser-adapter.ts 19.06
Prompt-injection corpus (page content ≠ control) libs/bellona/remote-protocol/src/browser-prompt-injection.ts 19.08
Unified secret/PII redaction engine libs/bellona/remote-protocol/src/redaction-engine.ts 27.05
Sensitive-path default-deny (secret stores) libs/bellona/remote-protocol/src/sensitive-paths.ts 27.03
Organization controls (recording, audit-export, providers) libs/bellona/remote-protocol/src/organization-control.ts 27.06
Org policy templates (review-only) + audit-export pinning libs/bellona/remote-protocol/src/organization-template.ts 27.06 / 27.11
Jurisdiction-aware retention libs/bellona/remote-protocol/src/retention.ts 27.16
Data residency (cross-border refusal) libs/bellona/remote-protocol/src/organization-governance.ts 27.08
Actor attribution (who, in the audit) libs/bellona/remote-protocol/src/actor-attribution.ts 27.13
Erasure / right-to-be-forgotten path libs/bellona/remote-protocol/src/erasure.ts —

Sign-off boundary (restated)#

This artifact prepares for, but does not constitute, legal sign-off. Before a deployment that captures any person's screen — and especially any deployment in an employment context — qualified counsel and the data-protection function must review: the lawful basis, the consent/notice model, the per-jurisdiction retention day counts, the cross-border transfer basis for any out-of-region storage, and any required employee-representative consultation. The [Counsel decision] markers above enumerate the specific questions to resolve. Engineering's role here is to have made the controls real and to have stated honestly what they do and do not cover; the legal determination is counsel's.