IRONWARD by AIFLOWLABS · In active development
When attacks coordinate, defence must too.
Governed AI security operations for Hong Kong teams without a security team.
IRONWARD’s AI agents investigate security issues and prepare the exact fix. Nothing changes until you, or a policy you approved, says yes. A separate check proves the fix worked and that your service still runs. Every step is kept as evidence.
A quiet architectural chamber. A single service core sits at the centre, with faint evidence paths resting around it.
02The problem
More alerts do not make a stronger defence.
A finding still needs someone to understand it, decide what matters, apply a fix and check the service still works. In a small team that work competes with everything else, so it waits.
HKCERT names AI-driven attacks and agentic AI risks first among its five key risks for 2026. As attackers automate, a stream of alerts is not enough. Defence has to follow through.
355,000 SMEs
98.5% of Hong Kong businesses.
15,877 incidents
handled by HKCERT in 2025, a record and up 27%.
Source · HKCERT Annual Report 2025
26% of SMEs
have dedicated security staff. Only 5% plan to hire.
Four disconnected evidence panels orbit the service, one for each unfinished step. One open work item stays unresolved.
03How it works
A swarm on your side, with one accountable workflow.
ObserveInvestigateProposeApproveActVerify
Specialist agents each have one job: investigate, prepare the response, or check the result. They share one task record instead of one long conversation.
- Find what matters. Turn observations into a focused investigation.
- Prepare the exact fix. The affected system, the change, its expected impact and how success will be checked.
- Prove the result. A separate check confirms the security outcome and that the business function still works. Missing evidence keeps the task open.
More agents is not the goal. Finished, verified work is.
Six stations form on two staggered arcs around the service, connected by evidence paths. Tool work is drawn as cubes, agent work as octahedra.
04Policy alignment
Built to the control model Hong Kong set for AI agents.
The 2026 Policy Address sets out how the Government itself runs AI agents. IRONWARD is designed around the same controls.
| Hong Kong’s control model (Policy Address 2026, footnote 12) | IRONWARD |
|---|---|
| “Human oversight should be maintained for high-risk actions” | No change without your approval or a policy you set in advance |
| “transmission of confidential or personal data to external platforms should be prohibited” | Models can run on your own premises |
| “risk assessments before deployment” | Every fix is proposed, scoped and assessed before it runs |
| “access control strictly enforced, the execution environment isolated” | Agents hold no credentials. A separate execution layer checks every action and runs it in isolation |
| “operation logs retained to ensure security, controllability and traceability” | A complete evidence record of every finding, decision, action and check |
Source: 2026 Policy Address, footnote 12. IRONWARD is designed to this model; Hong Kong’s AI Agent safety guidelines are due in 2027.
The chamber dims. Five paired lines of text face each other: the Government’s control on the left, IRONWARD’s mechanism on the right.
05Control and continuity
Agent speed. Human authority.
Operators approve changes one by one, or set bounded policies in advance for specific conditions and systems. The execution layer checks that each action matches the permitted system, operation and time window. Agents propose; they never create their own authority. Anything outside policy escalates to a person.
Stopping the threat is only half the job. IRONWARD checks the security outcome and the normal behaviour of the affected service. A proposed fix, an applied change and a verified outcome are different milestones, and you can see which is which.
Change: the proposed fix · System: the one named in the proposal · Window: 30 minutes
A translucent boundary stands between the agents and the service. It opens a narrow passage only for the approved change. A request outside policy is visibly refused.
06Your data, your infrastructure
Your data can stay on site.
IRONWARD is designed for locally hosted AI and approved cloud services. Routing follows your data policy, so you aren’t tied to one model vendor. Work records are kept separately from the agents, so interrupted tasks can be recovered and uncertain actions are checked before anything is retried.
Deployment options are released as each one is validated. We agree data handling, logs and support access with each partner before a pilot.
Evidence and prompts stay on your own infrastructure. If local AI is unavailable, a local-only task waits or fails openly. It never quietly switches to the cloud.
explains permitted routing · sends nothing
Four compute plinths stand in a ring inside a local boundary on the chamber floor. An optional path to an approved cloud provider leaves the boundary. The selector explains permitted routing; it sends nothing.
07Design partners
Start with one security workflow worth improving.
Who it’s for first: Hong Kong software and IT-service teams that own their code and run Linux services but have no security staff. That includes suppliers asked by banks, insurers or critical-infrastructure clients to prove their security.
The Security Diagnostic fee is credited to the subscription if you continue.
Three planes step forward in depth. The first, the Security Diagnostic, is highlighted as the entry point.
08Who is building it
Accountable people behind every result.
IRONWARD is owned and developed by AIFlow Labs Limited, a Hong Kong company. It is a separate product with its own codebase, not derived from client work. AI coding tools speed up engineering; named people are accountable for every result.
Igor Lessio
23+ years building software. Production multi-agent systems and data infrastructure; interim CTO for an AI data company. Leads architecture and engineering.
Hong Kong pilot engineer
Hong Kong-based. Leads pilot deployment and design-partner support.
Advisers
Named here once confirmed.
The chamber settles. Editorial typography only; no portrait is shown until one is supplied with permission.
09Progress, roadmap and commitments
Building the product. Proving the outcome.
Now: integrated components for task coordination, controlled execution, durable records and agent communication. Local testing has run a repository inspection with evaluation checks. The wider workflow is still being qualified.
Four architectural thresholds stand in a row receding into the chamber: now, next, then, and the open commitment.
Design partners
Put a swarm on your side.
Help shape governed AI defence around a real security problem. We’re looking for Hong Kong design partners, security practitioners and ecosystem partners willing to test the first workflows and challenge the results.
Stage
In active development. We are accepting Hong Kong design partners for paid, fixed-scope Security Diagnostics. General availability has not been announced.