Safety
Conventional agents
Added to each agent after it is built.
Mother Intelligence
Inherited by every child at creation.
AI safety, built for agents
Say yes to AI agents. Keep control.
AICE is building a connected safety and assurance layer for organisations putting AI agents to work: Agentic Defense Networks, the Agent Custodian, and SAFEAI architecture that brings them together.
Rolling out first to a small group of design partners through our Early Access Program, with hands-on deployment support from the aICE team.
Assurance layer oversight active
SAFEAI architecture
Governance · enforcement · evidence
Agent Custodian
Per-agent policy and oversight
Defense Networks
Estate-wide testing and defence
Independent Sentinels
outside the networkaICE deployment support
Designed for oversightScoped accessHuman checkpointsAuditable by design
The foundation
Mother Intelligence.
Safety from the start.
Instead of building agents one by one and adding safeguards later, a governing intelligence creates and directs purpose-built child agents. Scroll to see what changes.
Credentials
Conventional agents
Long-lived keys stored with each agent.
Mother Intelligence
Granted per task, then expired.
Memory
Conventional agents
Copied into each agent, hard to control.
Mother Intelligence
Held centrally; scoped to what the task needs.
Lifespan
Conventional agents
Runs indefinitely and may drift over time.
Mother Intelligence
Exists for one task, then is retired.
Accountability
Conventional agents
Logs scattered across tools and vendors.
Mother Intelligence
One coherent record of reasoning and actions.
Improvement
Conventional agents
Fix each agent separately.
Mother Intelligence
Improve a policy once; every child benefits.
Keep scrolling to explore the model
Mother Intelligence engine
SafetyCredentialsMemoryLifespanAccountabilityImprovement
Safeguards at birthMother Engine → Safety
One governing intelligenceChild agents · task-scoped
Two complementary controls
Protect each agent. Defend the whole estate.
AICE combines individual supervision with coordinated, continuous defence for the systems around your agents.
Per-agent supervision
Agent Custodian
A responsible guardian alongside each agent, keeping what it does within the bounds its owners intended.
- Verified identity, named human owner and agent inventory
- Least-privilege access granted per task and expired afterwards
- Behaviour checked against policy; pause when intent and actions diverge
- Human sign-off for irreversible, high-impact actions
- Tamper-evident audit trail and fleet-wide kill switch
Designed to work across vendors and frameworks, whether agents are built in-house or bought.
Estate-wide defence
Agentic Defense Networks
Coordinated teams of defensive agents that watch, test and protect AI systems around the clock.
- Detection agents watch traffic, tool calls and data flows for misuse or leakage
- Red-team agents probe deployed systems to help find weaknesses early
- Remediation agents propose or, when approved, apply fixes
- Participating organisations can share anonymised attack patterns
Every action is logged; high-impact changes can be held for human approval.
One architecture for trustworthy AI
SAFEAI connects controls into one assurance layer.
The Agent Custodian enforces controls on each agent. Agentic Defense Networks test and defend the whole estate. Independent Sentinels watch from outside the network, while aICE deployment support underpins the rollout.
SCROLL01 / 04
One architecture for trustworthy AI
SAFEAI connects controls into one assurance layer.
The Agent Custodian enforces controls on each agent. Agentic Defense Networks test and defend the whole estate. Independent Sentinels watch from outside the network, while aICE deployment support underpins the rollout.
Control the individual
Agent Custodian
Identity, permissions, policy checks, approvals and audit for every agent.
Protect the estate
Defense Networks
Continuous detection, authorised red-teaming and governed remediation.
Independent oversight
Sentinels
Separate-plane monitoring, escalation and a circuit breaker.
SAFEAI reference architecture — connects controls and evidence around the AI tools your organisation deploys.
Start with visibility
Understand your AI architecture before securing it.
You cannot protect what you cannot see. Every engagement starts with a living map of how AI actually runs across your organisation.
The output: an AI Architecture Risk Map
A ranked view of exposures, existing controls and gaps SAFEAI can help close—providing a baseline for the Custodian, Defense Networks and Sentinels.
LayerWhat we mapWhy it matters
ModelsHosted APIs, open-weight models, fine-tunes and embedded vendor AI.Each has distinct failure modes, data terms and update cycles.
Agents & toolsAgents, copilots, plugins, scripts and the tools they can call.Tool access is where an agent's mistake becomes a real-world action.
Data flowsWhat each agent can read, write, remember and send outside.See where sensitive data could leak or be poisoned.
Identities & permissionsWhich human or service each agent acts as, and with what rights.Over-broad permissions are a common source of risk.
Trust boundariesWhere agents meet the internet, other agents, partners and customers.Every boundary can expose systems to injection or impersonation.
Governed from request to result
One flow. Every action accounted for.
Each aICE-built agent task follows a governed path, so actions are planned, permitted, contained and checked before they reach the real world.
Task progress01 / 07
01
Governed stepsTask intake
Scroll to follow the task
Execution sequence7 checkpoints
01
Task intake
Record the request source, sensitivity and data it touches.
02
Plan
Mother Intelligence breaks the request into scoped steps and selects child agents, tools and data.
03
Policy check
The Agent Custodian checks the plan and grants scoped, time-limited access. Out-of-policy work is refused or escalated.
04
Execute
Child agents work in contained environments, with inputs and outputs checked for injection and data leakage.
05
Verify
Check results against the original goal. A failed check returns the task to planning.
06
Human sign-off
Payments, deletions, external messages and other high-impact actions wait for an approver.
07
Complete
Commit the approved action, seal the audit record and retire child agents and access.
Sentinels observe the full flow from outside the agent network
Independent by design
Sentinels watch the watchers.
An agent network should never be the only judge of its own behaviour. Sentinels sit on a separate oversight plane, outside the agents they monitor.
Separate infrastructure and credentials
No agent holds keys that reach the Sentinel plane.
One-way, read-only feed
They receive copies of logs, traffic, tool calls and traces—not instructions from agents.
They watch the watchers
Sentinels monitor the Custodian and Defense Networks as well as working agents.
Tamper-evident records
Observations go to an append-only store agents cannot alter.
Independent circuit breaker
A serious threat can trigger a halt to an agent, workflow or network.
TierTriggerResponse
ObserveNormal operationRecord and baseline behaviour
AlertUnusual activityNotify the response team with evidence
ContainLikely policy breach, leak or compromisePause the agent or workflow; revoke access
HaltActive attack or widespread failureCircuit-break the network; humans decide when to resume
From first pilot to full fleet
The right support at every stage.
Tools alone do not make an organisation safe. aICE works with your teams to put people, policies and process around AI adoption.
Readiness assessment
Discover AI tools in use, including unofficial ones, and map the data and systems they touch.
What you get
AI Architecture Risk Map, prioritised risk register and a 90-day plan.
Your team
CISO or security lead, IT and data protection officer
Secure pilot
Design a first use case on Mother Intelligence and SAFEAI, with the Custodian and Sentinels in place from day one.
What you get
A working, overseen agent and a pilot report.
Your team
Use-case owner, engineering and risk
Scale with assurance
Extend controls across teams and vendors; connect audit trails to security and compliance systems.
What you get
Organisation-wide AI policy, vendor checklist and connected audit trail.
Your team
Security operations, procurement and compliance
Ongoing assurance
Continue red-teaming and Sentinel monitoring, with periodic reviews and incident support.
What you get
Assurance reports and board or regulator briefings.
Your team
Executive sponsor and audit committee
Managed assurance
aICE operates Sentinels and Defense Networks with a named team.
Co-managed
Your security team runs daily monitoring; aICE supports red-teaming and escalations.
Self-managed with support
Run the platform in your environment with training, reviews and incident backup.
Every model includes staff training, agent incident playbooks and a clear escalation path. Standards and response-time commitments are agreed with each partner.
Building responsibly
Defensive tools must meet the same standard.
We apply SAFEAI to our own systems first: authorised testing, inspectable decisions, independent review and customer data protection.
- Testing only against verified, authorised systems
- Inspectable decisions from our defense agents
- Internal and external control reviews
- Anonymised threat intelligence sharing