— SECURITY · 07 OF 07 · AI SAFETY POLICY

Eight commitments on how we build, evaluate, and deploy AI.

Cyborgs are AI systems. Building AI systems for production work in real businesses comes with obligations — to customers, to end-users, to society. This page lists the commitments we hold ourselves to. Aligned with NIST AI RMF and EU AI Act risk-class framework.

FRAMEWORKNIST AI RMF + EU AI Act
RISK CLASSLimited (Annex III)
HUMAN OVERSIGHTRequired (T3/T4)
UPDATEDQuarterly

01The eight commitments.

01 · PURPOSE

Augment, don’t replace.

Cyborgs are designed to take repetitive work off humans, freeing humans for higher-order tasks. Customer requests that aim purely at headcount displacement without transition support fall outside our intended use.

02 · OVERSIGHT

Humans in the loop at every consequential action.

Our trust-tier system (T1–T4) defaults consequential actions to T3/T4 — human approval / never-alone. T1 autonomous status is granted per-action only after demonstrated reliability.

03 · TRANSPARENCY

Cyborgs identify themselves as such.

Every Cyborg-authored message in your tools (Slack, Teams, GitHub, etc.) is tagged [cyborg]. End-users always know they’re talking to a Cyborg. No deceptive impersonation.

04 · EVALUATION

Pre-deployment red-team + ongoing eval.

Each role-Cyborg passes through pre-deployment red-team for safety, jailbreak resistance, scope-bypass, and prompt-injection. Quarterly re-evaluations with third-party assessors (commencing Q4 2026).

05 · DISCLOSURE

Model + version visible in your dashboard.

Which underlying foundation model powers your Cyborg, and at what version, is visible in your dashboard. Material model swaps trigger 30-day prior notice. No silent upgrades or downgrades.

06 · PROHIBITED

Hard nos — in writing.

The AUP lists ten forbidden categories, including impersonation, deepfakes without consent, mass political messaging, and any criminal-grade use. Refusal is enforced at request time.

07 · INCIDENTS

Public post-mortems for harm-class incidents.

Where a Cyborg causes meaningful harm to a customer or end-user, we publish a post-mortem in /changelog within 5 working days — sanitised for privacy, but otherwise honest about cause and fix.

08 · FUTURE

Engaged with the regulatory frontier.

We track EU AI Act, India’s AI governance proposals, and US executive orders. Material regulatory changes drive product changes within the timeframe regulators set; we don’t lobby against safety floors.

02Risk classification.

Under the EU AI Act framework, Cyborgs are Limited Risk systems where the typical role is supportive (e.g., software engineering, customer support, marketing). They cross into High Risk only when configured for specific Annex III use cases (e.g., employment screening, credit scoring, critical infrastructure).

For High Risk configurations, additional controls apply: documented risk management system, conformity assessment, post-market monitoring, technical documentation per Annex IV. We will refuse High Risk configurations until we have those controls in place — not before, not even on customer demand.

03Foundation model posture.

We don’t train our own foundation models — we use providers (Anthropic, OpenAI, others as the market evolves). Our diligence on each provider:

  • Safety practices. Published safety evaluations, RLHF / Constitutional-AI training, jailbreak resistance scores.
  • Data handling. Contractual zero-retention of prompts. No training on customer data.
  • Reliability. Track record on production uptime, model deprecation policies, version pinning.
  • Geography. Where the inference compute is hosted, residency options.
  • Incident history. Public incident transparency, post-mortems, recall practices.

Provider changes that affect customer experience are surfaced to customers with notice. Our DPA sub-processor table is the source of truth on which providers are currently allowed.

04Bias + fairness.

Foundation models inherit biases from their training data. We don’t pretend otherwise. Our mitigations:

  • Role-shape constraints. Cyborgs operate within narrowly-scoped role contexts; bias surfaces less often than in open-domain chat.
  • Output review at trust tiers. T2/T3 actions get human review where bias would matter most (hiring, customer-facing decisions).
  • Customer-side eval hooks. Customers can configure custom evaluators that gate Cyborg actions on metrics they care about (sentiment, fairness, factual accuracy).
  • Annual bias audit. Internal review by a non-engineering team member, focused on customer-impacting behaviours. Public summary in the changelog.