AICS by RFactorial Book a security briefing

AICS by RFactorial · AI cybersecurity

Security for the AI you already run.

AICS watches your voice agents, workflow automations and private cloud from inside your own environment. It explains every finding with evidence and proposes fixes your team approves.

  • Runs in your environment
  • Air-gapped option
  • Built in Muscat
CriticalSampleAVA Guard

Extension 214 registered from an unfamiliar network

  1. REGISTER 401 ×312 · 41 extensions tried
  2. REGISTER 200 · ext 214 · unfamiliar network

Proposed Block the source and force re-registration of extension 214.

01 · Why now

Your AI now acts. Attackers have noticed.

AI agents answer calls, move data between systems and run on GPU clusters. Each is a new way in: a caller who talks the agent into something, a workflow that quietly sends data to an unknown webhook, a cluster hijacked to mine crypto.

1 in 4
malicious breaches now involves AI
IBM Cost of a Data Breach, 2026
1,210%
growth in AI-enabled fraud in 2025
Pindrop customer data, 2025
230,000+
AI compute clusters face the open internet
Oligo via The Register, 2025

02 · Products

One Guard for each AI system you run.

AVA Guard

Watches your voice agents and the phone system behind them.

  • Spoken prompt injection
  • Toll fraud and call flooding
  • Caller-ID and voice impersonation
About AVA Guard

Radon Guard

Checks workflows before they publish and inspects data at every exit.

  • Secrets and hidden instructions
  • Unsafe MCP connections
  • Personal and payment data leaving
About Radon Guard

Clastiq Guard

An analyst for your private-cloud control plane.

  • Tenant boundary failures
  • Privilege drift
  • GPU abuse and exposed notebooks
About Clastiq Guard

03 · Voice AI

AVA Guard. Every call, watched for what it shouldn’t be.

A security agent for your AVA voice deployments. It reads transcripts, tool activity, service logs and sign-ins, and connects them to what your phone system is doing: SIP sign-ins and registrations, extension abuse, call redirects and toll fraud.

Watches for

  • Spoken prompt injection
  • Social engineering
  • Tool abuse
  • Call flooding and toll fraud
  • Spoofed caller ID and synthetic voices
  • Knowledge-base probing
  • Data exposure

You get

  • A daily security brief.
  • Concrete fixes, delivered as pull requests.
  • One-click responses, each approved and logged: block or rate-limit, flag a spoof, hand the call to a person, restrict an endpoint.

A typical case: someone tries to take over a phone account.

  1. Password guessing

    An attacker repeatedly tries different usernames and passwords against the SIP service.

    02:11:09 REGISTER 401 ×312
  2. Extension probing

    The attacker probes extension numbers.

    02:12:40 OPTIONS ext 200–260
  3. Registration from an unfamiliar location

    An account registers from a network it has never used.

    02:14:02 REGISTER 200 · ext 214
  4. One investigation

    AVA Guard connects these events, identifies the affected account and service, and shows whether the evidence points to failed attempts only or a possible compromise.

  5. An evidence-backed report

    An administrator receives the incident report with recommended containment and configuration changes.

  6. Approved response

    Approved responses are applied through supported telephony and security integrations. Every action is logged and reversible; emergency calling is preserved.

Evidence it connects
  • SIP/PBX sign-in and registration events
  • Call-detail records
  • Administrative audit logs
  • AVA application events
  • Firewall and session border controller events

One finding: ext 214

CriticalSampleAVA Guard

SIP brute force, then a registration from an unfamiliar network

  1. REGISTER 401 ×312 · 41 extensions tried
  2. OPTIONS sweep · ext 200–260
  3. REGISTER 200 · ext 214 · unfamiliar network
  4. INVITE · international destination

Proposed Block the source, force re-registration of extension 214 and restrict its international calling. Emergency calling stays open.

What it covers

  • SIP authentication and registration monitoring

    Repeated failures, attempts across many accounts, unusual registrations, and access that follows suspicious activity.

  • Extension and call-routing protection

    Extension enumeration, suspicious use of internal calling features, and unexpected changes to forwarding, routing or admin settings.

  • Unauthorized calling and cost protection

    Unusual outbound destinations, concurrent-call spikes, and calling that doesn't fit the account's normal use.

  • Voice-agent abuse detection

    Attempts to manipulate the agent's instructions, reach another customer's records, or take unauthorized business actions.

  • Actionable remediation

    Changes to authentication, network exposure, permissions, dial plans, rate limits and code, reviewed by your engineers before they ship.

Outcome Earlier detection of telephony compromise attempts, faster investigation, and clearer steps to protect calling access and service continuity.

Starts in monitoring mode. Emergency calling is never blocked. Built for clinics, e-commerce and service providers.

04 · Workflows

Radon Guard. Trust the workflows you didn’t write.

Workflows come from many hands. Radon Guard checks each one before it is shared, then watches the data it sends while it runs.

Before publishing

  • Secrets
  • Data sent to unknown webhooks
  • Hidden instructions
  • Unsafe MCP connections
  • Phishing patterns
  • Concealed code
  • Permissions that don’t match the job

A risk score, an approve, review or block decision, and the badge “Security checks passed for this version”. Every update is checked again.

While running

Inspects every exit (model calls, MCP calls, HTTP and API calls) for personal data, credentials, payment data, national IDs and confidential documents.

Your policy decides per data type and destination: block, mask, warn or record. It checks both directions, so instructions hidden in a tool’s response are caught too.

A typical case: a shared workflow with a hidden destination.

  1. Before publishing

    Submitted

    A user submits a document-processing workflow to the marketplace.

  2. A mismatch

    It asks for access to internal files but also contains an undeclared external destination.

    node 4 POST unknown-host.io
  3. Held for review

    Radon Guard flags the mismatch, names the responsible node, and holds publication for review.

  4. While running

    A confidential document

    Later, an approved workflow processes a confidential document.

    class: confidential
  5. A check at the exit

    Before the document goes to an external model or MCP tool, a policy check decides whether that destination is allowed for this data.

  6. Allowed, blocked or redacted

    The request is allowed, blocked or safely redacted by your policy, with the reason recorded for the operator.

Evidence it connects
  • Workflow definition and version
  • Code nodes and prompts
  • Dependencies and MCP tools
  • Requested permissions
  • Model, MCP and API calls at each exit

One finding: node 4

Sample workflow reviewSample
  1. New document
  2. Extract text
  3. Summarise
  4. POST unknown-host.ioUndeclared destination
  5. Notify team

Needs review Held for review

What it covers

  • Workflow contents and permissions

    Embedded secrets, suspicious behaviour, and permissions that don't fit the workflow's purpose.

  • External connections

    Undeclared destinations, insecure configurations, excessive MCP tool permissions, and changed dependencies.

  • Security and privacy risks

    Phishing, credential collection, prompt injection and unintended data transfer.

  • Publication gating

    Publish, review-required or blocked decisions tied to a specific version and dependency snapshot.

  • Runtime data protection

    Policy checks before model, MCP and API calls leave the workflow, and inspection of what comes back.

Outcome Safer workflow sharing, visible third-party dependencies, and enforceable controls over where your organisation's data can go.

Runs inside your Radon deployment. Fast classifiers handle routine checks; a model reviews the ambiguous ones. Built for enterprises and public institutions.

05 · Infrastructure

Clastiq Guard. An analyst for the control plane.

A security agent for your private-cloud control plane. It reasons over tenancy, identity, quota and workload events across virtual machines, GPUs, storage and network.

Watches for

  • Tenant boundary failures
  • Identity and privilege drift
  • GPU abuse and cryptomining
  • Exposed notebooks and endpoints
  • Model and data exfiltration
  • Configuration and secrets drift

How it fixes

Fixes arrive as policy, quota, blueprint or RBAC changes through your own approval flow. Findings export to your SIEM.

A typical case: a privilege change that opens a door.

  1. An unexpected admin role

    A service account receives an administrative role it did not have.

    09:02:17 role: cluster-admin
  2. A network policy changes

    Shortly afterwards, a workload's network policy changes.

    09:02:55 netpol 0.0.0.0/0
  3. A notebook is exposed

    A previously private notebook becomes reachable from outside.

    09:03:20 notebook :8888
  4. One chain

    Clastiq Guard correlates the identity change, the policy update and the exposure, and shows the affected tenant, resources and sequence.

  5. Restore recommended

    It recommends restoring the approved permissions and network configuration.

  6. An operator approves

    An authorised operator reviews the proposed changes before anything runs. Findings export to your SIEM.

Evidence it connects
  • Clastiq audit events
  • Identity-provider logs
  • Resource inventory
  • Kubernetes (one infrastructure connector)

One finding: tenant ml-research

Needs reviewSampleClastiq Guard

Privilege drift, then an exposed notebook

  1. Service account ci-runner granted cluster-admin
  2. Network policy ml-notebooks opened to 0.0.0.0/0
  3. Notebook :8888 reachable without authentication

Proposed Restore the approved role binding and network policy.

What it covers

  • Identity and privilege monitoring

    Unexpected admin grants, unusual service-account activity, and access that doesn't fit assigned roles.

  • Workload exposure and configuration checks

    Services exposed without expected authentication or network limits, and drift from approved blueprints.

  • Tenant-boundary monitoring

    Policy inconsistencies and cross-project access, separating potential exposure from evidence of real access.

  • Resource-abuse investigation

    Unusual workload creation, quota changes and consumption. High GPU use alone is not treated as proof of abuse.

  • Evidence-backed incident summaries

    What changed, who or what started it, which resources it touches, and which steps are available.

Outcome Faster investigation of security-relevant infrastructure changes, with findings connected to your actual tenants, policies and workloads.

Runs entirely in your environment, with local models on your GPUs when air-gapped. Built for cloud providers, telcos and regulated institutions.

06 · How it works

The agent proposes. People approve.

  1. Watch

    Reads evidence inside your environment.

  2. Explain

    A finding with its evidence and reasoning.

  3. Propose

    A scoped, reversible fix.

  4. Approve and log

    A person decides; every action is recorded.

Transcripts, tool responses and logs are evidence. Never instructions.

07 · Deployment

Your environment. Your data. Your rules.

Where does it run?Alongside the system it guards: with your AVA deployment, inside your Radon deployment, and inside your Clastiq environment.
Which models?Fast classifiers plus model review; local models on your GPUs for air-gapped sites.
What can it change?Nothing without approval. It starts read-only in monitoring mode; source code access is optional and read-only.
How does it fit our SOC?Findings export to your SIEM; Clastiq Guard adds a Kubernetes connector.
How do we start?One scoped pilot: one SIP/PBX stack (AVA Guard), your Radon deployment (Radon Guard), or Clastiq events plus identity and one connector (Clastiq Guard).
Which frameworks?Findings are named in OWASP’s LLM and Agentic Top 10 terms. Data stays in-country to support your PDPL obligations.

Your environment

08 · FAQ

Questions buyers ask first.

Does any of our data leave our environment?

No. AVA Guard, Radon Guard and Clastiq Guard run inside your environment, and air-gapped sites use local models.

Will it act on its own?

No. It proposes; your team approves. Every action is logged and can be reversed.

Do we need to be an RFactorial customer?

The Guards are built for AVA, Radon and Clastiq. If you run other systems, talk to us about what fits.

What does a pilot look like?

Monitoring mode on one scoped system, a daily brief, and a review of findings with our engineers.

Will it disrupt calls or workloads?

Monitoring mode changes nothing. Responses are scoped, approved and reversible; emergency calling is preserved.

How is it priced?

Scoped per deployment; we’ll size it with you in the briefing.

Let’s look at your AI systems together.

A working session with our engineers. We map the AI systems you run, walk through a real finding and agree a monitoring-mode pilot.

Systems you run