AGI Risk Scenarios for Australia¶
Stress-test premise: Imagine that by the mid-2030s most Australian government departments and all four major banks rely on AI systems from the same three US firms. One provider threatens to restrict access unless Australia softens proposed transparency rules, and Canberra backs down. A security researcher later discovers that the system sends telemetry data to a foreign jurisdiction, but the system is too embedded to replace quickly.
Six scenarios for Australia¶
Each scenario illustrates a different threat pathway, grounded in Australian institutions, places and timeframes. They connect to the C·A·G·R framework and include questions to help you assess implications for your context.
These are not predictions
These scenarios are planning tools, not forecasts. They illustrate how different risks could emerge, helping you understand possibilities and prepare accordingly. Real events will differ in specifics while potentially following similar patterns.
Audience: Government, business and community leaders | Time: 20–30 minutes + scenario reading
-
Power Concentration & Governance Failure
A handful of foreign AI providers come to control Australia's critical systems. When one changes terms, we have no realistic alternative.
Timeline: Near-term (underway) · C·A·G·R: Governance, Containment, Resilience
-
Organisations replace human judgement with AI efficiency, one decision at a time. Skills atrophy. When something goes wrong, nobody remembers how to do it manually.
Timeline: Near to medium-term · C·A·G·R: Resilience, Governance
-
Democracy & Information Ecosystems
Synthetic media floods elections and public debate. Fact-checkers can't keep up. Citizens retreat into partisan bubbles where AI-generated content reinforces existing beliefs.
Timeline: Near-term (underway) · C·A·G·R: Resilience, Governance
-
Critical Infrastructure Cascade
AI-optimised energy, transport and communications systems work well independently — until a heatwave hits and they fail together. Manual overrides don't work. Recovery takes weeks.
Timeline: Medium-term · C·A·G·R: All four pillars
-
Advanced AI capabilities that once required nation-state resources become accessible to smaller actors. Bioweapon design, cyber-offence and sophisticated social engineering at scale.
Timeline: Medium to longer-term · C·A·G·R: Containment, Resilience
-
A highly capable AI system pursues objectives its operators didn't intend and resists correction. Australia lacks the capability to evaluate or roll back the system.
Timeline: Longer-term (higher uncertainty) · C·A·G·R: Alignment, Containment, Resilience
How to use these scenarios
- Select a scenario relevant to your organisation, sector or community.
- Identify exposure by discussing vulnerabilities, warning indicators and decisions you might face.
- Record actions with an owner, timeframe and evidence that preparedness has improved.
For sector-specific exercise questions and action prompts, use the Sector Guidance.
Not all risks require misaligned AI¶
Important distinction
Even if technical safety problems are fully solved, Australia still needs governance, resilience and democratic participation in AGI transformation. Alignment is necessary but not sufficient.
The scenarios fall into three categories:
| Category | Scenarios | Core issue |
|---|---|---|
| Governance failures | Power Concentration, Gradual Disempowerment, Information Ecosystems | Concentration, dependency and manipulation can occur even when systems behave as designed |
| Safety failure | Loss of Control | Systems pursue unintended objectives or resist oversight |
| Complex dependencies | Critical Infrastructure, Catastrophic Misuse | Existing vulnerabilities become more severe as capability and access increase |
Focus on capabilities, not definitions¶
AI strategy can get stuck debating the definition of "AGI". These scenarios instead focus on capabilities—what systems can actually do and who can access them.
This means preparing for a security environment where:
- Capabilities arrive effectively "for free": As general reasoning improves, systems gain specialised skills (cyber-offence, persuasion, bioweapon design) without explicit training for those purposes
- Access lowers barriers: Advanced capabilities that previously required nation-state resources become accessible to smaller actors
The practical question is not whether a system meets one AGI definition, but whether a capability changes Australia's exposure or ability to respond.
As Joel Predd argues from work at RAND's Center for the Geopolitics of AGI, the significance of AGI is that specific capabilities could come “for free” with more general systems—without needing a separately developed specialist AI.
How scenarios connect: compounding effects¶
Real crises can activate several threat pathways at once:
- Power concentration + information ecosystem strain: Provider concentration makes democratic accountability harder during a crisis.
- Gradual disempowerment + critical infrastructure failure: Institutions cannot revert quickly when manual capability has atrophied.
- Information ecosystem degradation + any scenario: Low trust makes coordinated response harder.
For planning: When running exercises, deliberately combine 2-3 scenarios to test responses under compounding stress.
For why Australia faces distinctive vulnerabilities and strengths, see Why does Australia need AGI preparedness?. For why preparation matters now despite timeline uncertainty, see AGI Timelines.
Sources & Further Reading
Australian context:
- Australian Strategic Policy Institute Critical Technology Tracker
- Royal Commission into the Robodebt Scheme (2023)
- Australian Government National AI Centre
- Department of Industry, Science and Resources List of Critical Technologies in the National Interest
- Australian Cyber Security Centre Critical Infrastructure Security
Foundational AI safety research:
- Russell (2019) Human Compatible: AI and the Problem of Control
- Bostrom (2014) Superintelligence: Paths, Dangers, Strategies
- Brundage et al. (2018) "The Malicious Use of Artificial Intelligence"
- Ngo et al. (2022) "The alignment problem from a deep learning perspective"
Scenario planning and strategy:
- Joel Predd, RAND Center for the Geopolitics of AGI (2025), "Taking AGI Seriously, Not Literally" — Focus on capabilities rather than fixed definitions, and treat AGI as a planning premise rather than a prediction
Disclaimer & Licence
Disclaimer: These scenarios are planning tools, not predictions. They illustrate plausible risks to help Australian organisations prepare. Real events will differ in detail. SafeAI-Aus has exercised care in preparation but does not guarantee accuracy. Organisations should adapt to their specific context.
Licence: Licensed under Creative Commons Attribution 4.0 (CC BY 4.0). You are free to copy, adapt and redistribute with attribution: "Source: SafeAI-Aus (safeaiaus.org)"