Draft — not legal advice. Lines marked [PLACEHOLDER] or [CONFIRM] are unfinished. Do not treat them as facts.

Trust and security

DRAFT. Every [PLACEHOLDER: ...] and [CONFIRM: ...] must be verified before it is treated as a fact. Do not describe a control we cannot evidence. This is the trust page, published at /security.

Tabletop exercises show how a leadership team would respond to a ransomware attack, a breach or a fraud. That information is sensitive. This page explains how Critios protects it, as far as the application code supports the statement.

Questions or security review requests: security@cbilatam.la


At a glance

  • Sealed answers stay off other participants' screens. A participant can read their own submission. Other participants can read a turn's submissions only after that turn is revealed. The rule is Postgres row-level security, not only the screen. The facilitator's session is allowed to read every submission; the facilitator console does not render the choice or the rationale until the reveal. [CONFIRM: whether that facilitator read, before the reveal, is what you want to promise. The rows are loaded into the facilitator's browser.]
  • Private notes stay with the facilitator. They are stored apart from the exercise state. Row-level security lets only the facilitator read them. Participants do not receive them. The AI code does not read them. They can appear in the facilitator's own report.
  • AI is requested, not automatic. A model is called only when a facilitator asks for a risk summary, a debrief narrative or a scenario draft. AI drafts need facilitator approval before participants see them. A drafted scenario cannot be played until it is saved.
  • The demo stores nothing you type. /demo keeps its state in the browser and does not write it to the database.
  • The pilot request form asks for name, email, company and an optional message. Analytics and the LinkedIn Insight Tag are described in the Privacy Policy. Google Analytics 4 is implemented (Consent Mode v2, production host only, Google ad storage always denied, paths redacted). The LinkedIn tag loads only on that same consent choice, and only when a partner ID is configured. [CONFIRM: GA4 property settings, and that the LinkedIn IDs in the deployment match this page.]
  • No overstated certifications. Critios maps your exercise results to SOC 2, NIST CSF 2.0, ISO/IEC 27001, PCI DSS v4.0, NIS2 and DORA. That is not a statement that Critios itself is certified. See compliance status.

Data handling

DataWhere it comes fromWhat we do with it
Pilot request (name, email, company, optional message, page language, landing version)Website formUsed to reply and arrange the session. The hidden anti-spam field is not stored. Kept for up to 6 months, after which it is deleted.
Demo inputs/demoDummy data in the browser; not saved to the database.
Facilitator accountMagic-link sign-inEmail address, via Supabase Auth. [CONFIRM: which system actually sends the message.]
ParticipantRoom codeDisplay name and role, on an anonymous session. No participant email is collected. [PLACEHOLDER: confirm if further fields are added.]
Decisions and sealed answersParticipants during exercisesStored for the debrief and the report. Other participants cannot read them until the reveal. Not included in the AI prompt until revealed.
Private notesFacilitatorFacilitator-only. Not sent to participants or to the model.
Scores, debriefs, reportsGenerated by the ServiceAvailable to the facilitator; the report can be downloaded as a PDF.
LogsPlatform operationSecurity and troubleshooting. Retained for [PLACEHOLDER: retention period].
  • Data location: [PLACEHOLDER: hosting region or regions, e.g. US / EU].
  • Retention and deletion: Pilot request submissions and exercise data (exercises, decisions, notes, reports and AI drafts) are kept for up to 6 months, after which they are deleted. [PLACEHOLDER: deletion window after a contract ends, and whether remaining account data is deleted or returned.] Backups age out within [PLACEHOLDER: backup retention period].
  • Customer ownership: Customers own their exercise data. See our Terms and Privacy Policy.
  • Recommendation: Exercises don't need real secrets. Don't enter real credentials, card data or personal data about real incident victims.

How sealed answers work

  1. Each participant submits a decision privately, against a timer.
  2. Until the facilitator reveals the turn, other participants cannot read that turn's submissions. The submitting participant can read their own. The facilitator's database access can read them throughout; the console hides the contents until the reveal. [CONFIRM: product promise for facilitator access before reveal.]
  3. When the turn is revealed, the answers become part of the shared debrief.
  4. Private notes are not revealed to participants. Only the facilitator can read facilitator_private. [CONFIRM: staff with production database access can also read them. State who that is, and how access is approved and logged.]
  5. The copilot prompt is built by buildContext, which keeps only revealed turns. Private notes are not queried. [CONFIRM: publish this as the enforcement point. Unrevealed rows are read on the server and then omitted from the prompt; they are not sent to the model.]

AI use

  • What the exercise copilot sees: revealed decisions (role, question, choice, rationale), the scenario title, family, summary, objectives and role titles, the chaos level, scores, inject titles and situations for revealed turns, applied outcomes, and up to 60 room-visible events with names from join events removed.
  • What a scenario draft sees: the facilitator's brief and Critios's trigger and control catalogs. Not an exercise record.
  • What it does not see: unrevealed submissions, and private notes.
  • No request, no call. Running an exercise does not by itself send exercise content to a model.
  • Model path: both calls use the Vercel AI SDK with model id anthropic/claude-sonnet-5.5 (Vercel AI Gateway, then Anthropic). Each call sets gateway zeroDataRetention: true. [CONFIRM: the gateway and Anthropic honor that flag, and no production integration bypasses the gateway.]
  • Training: [CONFIRM: no training on customer data — Critios does not use Customer Data to train models, and the Vercel and Anthropic terms that apply to these calls prohibit training on that data.]
  • Human approval: risk summaries and narratives are stored in Supabase and are not shared with participants until the facilitator approves them. A risk note is shared only by a further action. An AI scenario stays a draft until the facilitator saves it in the editor.
  • Human oversight: AI output is facilitation support. Review scores and reports before you rely on them.

Encryption and transport

  • In transit: [CONFIRM: TLS version at the Vercel edge for www.critios.com, and whether HSTS is enabled. The application itself does not set a Strict-Transport-Security header.]
  • Headers the application does set: Content-Security-Policy of frame-ancestors 'self', object-src 'none', base-uri 'self', form-action 'self'; X-Frame-Options: SAMEORIGIN; X-Content-Type-Options: nosniff; Referrer-Policy: strict-origin-when-cross-origin; a Permissions-Policy that disables camera, microphone, geolocation, payment and USB; X-Powered-By removed. The content-security policy does not yet restrict scripts. [CONFIRM: anything the edge adds on top.]
  • At rest: [CONFIRM: database and backups encrypted at rest — algorithm and who provides it.]
  • Secrets: the service-role key and email API key are read from environment variables, not from the repository. [CONFIRM: where those variables live (Vercel project settings) and who can read them.]

Access control

  • Customer access: facilitators sign in with an email magic link (Supabase). Participants use an anonymous session plus the room code. [CONFIRM: no password, no MFA and no SSO in the application today — say so, or update this when they exist.] Roles in an exercise are the facilitator and the seats (Chief Executive, CISO, CFO, General Counsel, Head of Comms, or the scenario's roles).
  • Tenant isolation: [CONFIRM: row-level security. Participants read their own unrevealed submissions and other participants' submissions only after reveal. Only the facilitator can read the exercise row, private notes and AI drafts. Writes from the browser are denied; changes go through server actions. State this only after the policies are reviewed as the published control.]
  • Critios staff access: [CONFIRM: production access limited to named personnel, MFA required, least privilege, access logged.] Staff should access Customer Data only to provide support you request, to keep the Service secure, or when the law requires it. The server uses a service-role key that bypasses row-level security; authorization is checked in the server action. [CONFIRM: who can use that key.]
  • Logging: exercise actions are appended to an event log (room-visible or facilitator-only). [CONFIRM: a separate audit of staff access to production data.]

Infrastructure and operations

  • Application: Next.js, served at www.critios.com. critios.com redirects there.
  • Hosting: Vercel (application and AI Gateway), Supabase (database and authentication). [PLACEHOLDER: regions for Vercel and Supabase.]
  • Email: Resend is used by the application to notify the team of a pilot request, when RESEND_API_KEY and LEAD_NOTIFY_EMAIL are set. [CONFIRM: Resend's legal entity, and whether magic-link email also goes through Resend.]
  • Backups: [CONFIRM: backup frequency, encryption, restore testing.]
  • Secure development: [CONFIRM: code review, dependency scanning, and whether production data is used in development. Dev and tests currently share the production database — do not describe an isolated staging environment until one exists.]
  • Incident response: [CONFIRM: the incident response process.] We will notify affected Customers of a security incident affecting their data without undue delay and within [PLACEHOLDER: notification commitment, e.g. 72 hours] of confirmation.
  • Penetration testing: [PLACEHOLDER: last independent penetration test date and provider, or "not yet performed".]

Subprocessors

Third parties that may process Customer or website data on our behalf. We will give [PLACEHOLDER: notice period, e.g. 30 days] notice before adding or replacing a subprocessor that processes Customer Data. [PLACEHOLDER: how customers subscribe to that notice.]

SubprocessorPurposeData processedLocation
Vercel Inc.Application hosting, Web Analytics, AI GatewayRequest logs; cookieless page analytics if the product is enabled; AI prompts (revealed decisions and scenario context, or a scenario brief)[PLACEHOLDER: region]
Anthropic PBC (via Vercel AI Gateway)Model for the copilot and scenario drafts, only when a facilitator requests oneThe prompt described above. Not sealed answers or private notes[PLACEHOLDER: region]
Supabase Inc.Database and authentication (magic link for facilitators, anonymous sessions for participants)Account email, exercise data, AI drafts, pilot request submissions[PLACEHOLDER: region]
Resend [CONFIRM: legal entity name]Email notification of a pilot request, when configuredName, email, company, message[PLACEHOLDER: region]
Google LLC (Analytics 4 via gtag.js, not Tag Manager)Product and site analytics. Consent Mode v2. Google ad storage denied. Paths redacted. Production host onlyCookie ids, redacted page paths, allow-listed event parameters. Not exercise answers or notes[PLACEHOLDER: region] [CONFIRM: property settings in GA admin]
LinkedIn (Insight Tag)Campaign measurement, only if a partner ID is set, on the production host, after the same consent choice as analyticsLinkedIn's cookies and page views. The conversion event is an id only[PLACEHOLDER: region, partner ID, conversion ID]

No separate form or CRM product is called by the application. [PLACEHOLDER: add one here if it is introduced.] No error-monitoring product is wired in the repository. [PLACEHOLDER: add Sentry or similar here if it is enabled outside the repo.]

Data Processing Agreement

A DPA covering GDPR Art. 28 and Argentina Law 25.326 is [PLACEHOLDER: "available on request at the privacy contact email" / "linked here" / "in preparation"]. International transfers are covered by [PLACEHOLDER: transfer mechanism, e.g. EU SCCs].

Compliance status

Critios helps you produce evidence for frameworks. Critios itself is not currently certified or attested against them unless a row below says so. None of these rows should say "certified" without a report.

FrameworkCritios's own status
SOC 2[PLACEHOLDER: e.g. "Not audited" / "Type I in progress, target date" — do not state certified unless a report has been issued]
ISO/IEC 27001[PLACEHOLDER: e.g. "Not certified" / "Planned"]
GDPRWe aim to comply as controller (website and pilot request form) and processor (the Service). [PLACEHOLDER: counsel to confirm this wording]
Security questionnairesWe can complete standard questionnaires (for example [PLACEHOLDER: CAIQ / SIG Lite / custom]) on request.

Vulnerability disclosure

We welcome reports from security researchers.

  • Report to: security@cbilatam.la ([PLACEHOLDER: PGP key URL, or remove]). Machine-readable details: https://www.critios.com/.well-known/security.txt.
  • Include: affected URL or component, steps to reproduce, impact, and your contact details.
  • Our commitment: acknowledge within [PLACEHOLDER: e.g. 3 business days], give an initial assessment within [PLACEHOLDER: e.g. 10 business days], keep you informed, and credit you if you wish.
  • In scope: www.critios.com and critios.com (which redirects to www). [PLACEHOLDER: other hostnames, if any. The application is not on a separate app subdomain.]
  • Please don't: access or modify other users' data; run denial-of-service, spam or social-engineering tests; use automated scanners at high volume; or disclose publicly before we have fixed the issue (or [PLACEHOLDER: e.g. 90] days have passed).
  • Safe harbor: We will not pursue legal action against good-faith research that follows this policy. [PLACEHOLDER: counsel to confirm this safe-harbor wording.]
  • Bounty: [PLACEHOLDER: "We don't currently offer a paid bug bounty" or the bounty terms.]

Last updated: [PLACEHOLDER: publication date]

Run this exercise with your team.

Run it yourself on Critios: US$490 per exercise, no subscription. The self-serve platform opens soon.