IN
Untrusted inputs become instructions.
Direct prompts, documents, webpages, emails, audio and retrieved content may compete with the organisation’s intended rules.
Independent assurance for Australian organisations
Independent AI Security Testing and Assurance
Phase Security tests whether AI systems can be manipulated, disclose protected information or take unsafe actions—and helps organisations establish the governance, trust boundaries and operational controls needed to use AI responsibly.
IndependentVendor-neutral assessment
AuthorisedControlled testing only
Evidence ledFindings, severity and remediation
Human accountableHuman-reviewed conclusions
The AI trust boundary
AI creates new paths between untrusted input and real-world action.
External models, embedded content, retrieval systems, memory, voice interfaces and connected tools can all influence an AI system. The question is not only what the model says—but what it can reach, reveal and cause.
ACT
Agents turn output into action.
When models can search, send, modify, approve or transact, a persuasive response can become an operational event.
GOV
Governance needs technical evidence.
Policies are only reliable when ownership, access, escalation, human intervention and control effectiveness can be demonstrated.
Phase maps the trust boundary, tests the plausible paths through it and turns the evidence into practical controls, accountable decisions and a retest plan.
AI security services
Start with exposure. Go deeper where the evidence points.
Each engagement is scoped to the system, decision or use case in front of us. The services can stand alone or form a staged assurance program.
AI Security Exposure Review
Establish where AI is used, what it can access, how external models enter the operating environment and where control could be lost.
- AI system and use-case inventory
- Data and integration exposure map
- Trust-boundary and ownership gaps
- Prioritised next actions
Prompt Injection & Agent Security Assessment
Controlled adversarial testing of text and voice systems, retrieval, memory, agent permissions and connected tools.
- Direct and indirect injection tests
- Text and voice attack paths
- RAG, memory and tool-use scenarios
- Evidence, severity and retest plan
AI Governance & Operational Assurance
Connect policy to the way AI is selected, approved, monitored, challenged, escalated and recovered in practice.
- Accountability and approval controls
- Human review and override
- Supplier and external-model governance
- Incident and continuity readiness
01
Scope and authorisation
Define the system, business purpose, permitted test surfaces, safety limits, contacts and stop conditions.
02
System and trust-boundary mapping
Map models, interfaces, data sources, retrieval, memory, identities, permissions, tools and human decision points.
03
Baseline test design
Create repeatable text and voice scenarios against defined control objectives, expected behaviour and evidence requirements.
04
Controlled execution
Test direct and indirect prompt injection, data disclosure, instruction conflict, unsafe tool use and persistence within the agreed boundary.
05
Evidence and severity
Record inputs, outputs, system state, reproducibility, business consequence and the controls that succeeded or failed.
06
Remediation and retest
Prioritise containment and design changes, assign owners and define the conditions for verifying improvement.
Operational foundation
Security across the whole operating environment.
Phase Security draws on more than two decades of leadership across facilities management, physical security and electronic security operations, combined with focused expertise in cybersecurity, AI security, governance, risk and continuity.
That breadth matters because modern security failures do not respect organisational boundaries. People, premises, identities, cloud services, external models and automated decisions increasingly affect one another. Phase examines those connections, tests where control can be lost and turns the evidence into practical action.
Assessment evidence
A practical record of what was tested and what happens next.
Deliverables are tailored to scope. Findings distinguish observed evidence from assumptions and immediate exposure from longer-term improvement.
- Executive exposure summary
- AI system and use-case inventory
- Trust-boundary and attack-path map
- Controlled test evidence
- Risk-rated findings
- Immediate containment actions
- Prioritised 30/60/90-day roadmap
- Named control owners
- Remediation and retest plan
- Specialist referral requirements
The purpose is not to produce a theatrical “hack.” It is to establish whether a meaningful failure can be reproduced, what consequence it creates and which control should change.
Phase Lab · Research with status
Build evidence. Preserve the lineage. Retire what no longer fits.
Phase Lab keeps exploratory work separate from customer-ready capability. Active research is named plainly; earlier ideas may be absorbed, paused or abandoned without being rewritten as finished products.
Prompt Injection Testing
A repeatable testing and reporting method for text and voice systems, including indirect content, retrieval, memory, agents and tools.
AI Trust Boundaries
How internal systems gain value from external and embedded models without surrendering control of protected data, decisions or actions.
HaaD Defensive Deception
Controlled synthetic substitution intended to shield authoritative truth, increase adversarial uncertainty and produce useful defensive evidence.
Research lineage
This is a record of thinking, not a catalogue of products. It shows how earlier questions shaped the narrower work Phase is pursuing now.
CORALISAbsorbed
Early Phase work on readiness, risk, maturity, oversight and evidence—conceived before the Australian Government established its Office of AI in July 2026. Its useful logic now informs the AI Security Exposure Review and AI Governance & Operational Assurance rather than being marketed as a separate framework.
HaaS → HaaDEvolved
HaaS framed hallucination and synthetic model behaviour as a security surface. HaaD opened a distinct research line: whether controlled synthetic responses could be repurposed as governed defensive deception. The concepts remain separate; HaaD is the active demonstrator.
Other conceptsPaused
AFA, AQDA, CrypCycle, AI Resistance and post-quantum concepts remain in the internal research record. They are not presented here as current services or production-ready capability.
Research rule: public claims follow current evidence. Synthetic defensive material must never contaminate an authoritative source of truth.
About Phase Security
Independent, operational and deliberately bounded.
Phase Security is a Sydney-based AI-security and operational-assurance practice. It focuses on prompt injection, agent security, AI trust boundaries, governance, risk and continuity where they affect sensitive information, critical decisions or real-world operations.
Cyber and continuity work remains part of the picture where it supports AI security, critical operations or recovery. Phase is not positioned as a general construction, IT support or facilities-consulting business.
AIeOU Pty Ltd is the research and portfolio authority behind Phase Lab. That relationship allows ideas to be explored and recorded without presenting unfinished work as customer-ready capability.
Know what your AI can trust—and what it can do.
Begin with a short fit check. Phase will confirm the system or use case, the concern, whether controlled testing is appropriate and what evidence would be required.