Skip to content
InovagioInovagio

GuardAI safety

Content safety on every prompt, response and tool call.

Screens every prompt, response and tool call against the limits you set, works offline, and blocks when in doubt.

We’ll spend 30 minutes with you on one of your workflows, and you don’t need to share any data.

The Inovagio App at step 4 for the example company Harlow Distribution: the privacy and security lead approves how each kind of data is protected. Personal and card data are labeled field by field, bank details are masked on every card but the bank-detail check and never sent to a model, and tax IDs show only their last four digits.

What Guard does

Guard is Inovagio’s AI safety layer: it screens every prompt, model response and agent tool call against per-category policies you set, runs offline by default, and fails closed.

  • Per-category thresholds

    Set a 0 to 100 threshold for each harm category, separately for prompts and responses, per tenant or per request.

  • Requests can only tighten policy

    An operator floor caps every tenant and request setting, and an unset policy resolves to strict.

  • Offline by default

    Built-in analyzers run inside your deployment; external moderation services are a per-tenant opt-in, never enabled in government, health or finance profiles.

  • Injection defenses

    Instructions found in documents, email or tool output never carry authority: an agent’s permissions come only from its grant, so an injected request still needs an allowed tool and, where policy says, a person. Guard also detects injection and jailbreak attempts, scans tool and document content, and removes injected control tokens before tokenization.

  • Streaming enforced

    Streamed output is checked segment by segment through a bounded hold-back buffer, and a blocked stream ends with a clear safety reason.

  • Tool calls gated

    The agent loop checks every tool call’s input before it runs, and a blocked input becomes a recorded, audited refusal.

  • CSAM always blocked

    A non-adjustable classifier on text and images that a production deployment refuses to start without, plus matching against industry hash lists where your organization is licensed to use them.

  • Enforces the kill switch

    When Archon’s kill switch is pulled, Guard halts one Virtual Employee, a department or everything at once within seconds, with health and admin routes kept up.

  • Privacy-safe audit

    Every decision records category, severity, threshold and action; personal data it finds is recorded as category and count only, and retention defaults to one year.

Guard at work

In the App at step 4: Check, protect and prove the data.

  • Personal data and card data, labeled field by field in every source.
  • Bank details masked on every card but the bank-detail check, and never sent to a model.
  • Tax IDs show only their last four digits.
The Inovagio App at step 4 for the example company Harlow Distribution: the privacy and security lead approves how each kind of data is protected. Personal and card data are labeled field by field, bank details are masked on every card but the bank-detail check and never sent to a model, and tax IDs show only their last four digits.
Each kind of data found gets its protections, approved by privacy and security before any model reads it.

Inside Guard

The views Guard gives the people who run it, on sample data from Inovagio’s example companies.

The Security page in the Inovagio App for Harlow Distribution, an example company with example data. It shows the quarter’s access reviews and Guard’s screening for the last seven days, by kind: Guard masks personal data, refuses hidden instructions and unsafe content, and holds bank details sent by email for a person.

Guard › Overview: Guard’s screening for the last seven days, by kind, with what was masked, refused or held for a person. Sample data from Inovagio’s example companies.

Everything Guard does, by area.

Categories

  • Sexual content, violence, hate, self-harm
  • Profanity and political content
  • Personal and health information
  • Custom blocklists
  • Prompt injection and jailbreaks
  • Dangerous instructions, illicit drug manufacture, malware and cyber-attack instructions
  • CSAM (mandatory hard block)

Where it enforces

  • The inference serving path
  • The gateway, including requests routed to external models
  • Agent tool-call inputs
  • Streaming output for console, CLI, SDK and desktop
  • Agent and Virtual Employee operations through Guard hooks

Policy and control

  • One policy schema in three layers: operator, tenant, request
  • Annotate-only and annotate-and-filter modes
  • Policy writes authorized per principal
  • Versioned policy snapshots held by every in-process filter
  • Guard service for policy management and on-demand scans
  • Enforces the kill switch for one Virtual Employee, a department or everything
  • Live policy reload without restarting sessions

Assurance

  • Fail closed on any analyzer error
  • Red-team tests as an acceptance gate
  • Blocking tuned per category to keep false positives low
  • Salted hash fingerprints, never for personal data
  • Audit retention configurable to the period your regulator and records schedule require

What Guard replaces

Guard does the work of these, inside one platform with one identity, one policy and one record:

  • Cloud content-moderation services
  • Stand-alone prompt-injection filters
  • Hand-maintained keyword blocklists
  • Ad hoc kill switches for AI agents

The controls Guard brings

Every Inovagio product keeps the same five commitments, because regulated work needs all five at once.

  • Security

    Prompt injection, personal data and unsafe content are screened on the way in and out.

  • Governance

    Safety rules are set per tenant, workflow and model.

  • On the record

    Every block and redaction is recorded with its reason.

  • Monitoring

    Safety events appear in Panoptes with the rest of the platform.

  • Cost per task

    Checks run on your hardware, with no per-call charge to an outside service.

Guard in your industry

Every industry runs the same ten steps, so Guard does the same job in each, on that industry’s own work.

For developers: Guard at a glance

Response annotations
Safety results in the Inovagio API’s response fields; a clear error for a blocked prompt
Modes
Annotate-only; annotate-and-filter
Analyzers
Built-in offline analyzers; offline CSAM classifier (mandatory) and hash matching where licensed; optional external moderation (opt-in); local Inovagio safety classifier for semantic categories
Default posture
Strict, fail-closed, offline
Audit retention
1 year by default, extendable to the retention period your regulator and records schedule require
Runtime
In-process filters on every path, plus a supervised Guard service
Platforms
macOS, Linux and Windows servers

Request the developer editionSee the architecture

Guard in the platform.

Where Guard fits

One platform
  1. Agent HarnessAgent harness

    Inovagio CLIInovagio DesktopWeb ChatInovagio SDK

  2. OpsisExperience layer

    FolioMorphe

  3. Virtual EmployeesVirtual Employee layer

    NexusErgon

  4. SynesisKnowledge layer

    PerceptionMNEMOS

  5. OriginModel layer

    Inference ServerForge

  6. Phylax (this layer)Trust layer

    AuthInovagio SecretsGuard (this product)

  7. OperationsOperations

    PanoptesArchon

  8. Platform servicesPlatform services

    ServerAPI ManagementConnect

Where Guard fits

Further reading

Guides and posts on Guard

See Guard on one of your workflows.

Every project starts with one workflow your people know well, measured against what it costs you today.