Chapter 01

Risk map

Filter by vector, evidence level, severity and horizon. Each entry opens the full chain: precondition, trigger, cascade and impact.

Vector

Evidence

Severity

Horizon

Figure 1.1 · Risks by horizon and severity. 33 of 33 risks.

Already happening risks with harm already documented

1–3 years plausible materialisation in the short term

3–10 years depend on capabilities that do not exist yet

Military and autonomous weapons · Political and power concentration

Arms race between states

Military competition over AI capabilities pushes deployment before reliability is known, and the forum that could slow it operates by consensus among the competitors.

Severity: Catastrophic
Horizon: Already happening
Evidence: Observed
Consensus: Medium

Chain of materialisation

PreconditionObserved
Programmes are running and contracts signed
In July 2025 the Department of Defense awarded contracts of up to 200 million dollars each to Anthropic, Google, OpenAI and xAI. Replicator was announced to field thousands of uncrewed systems by August 2025; Congress records that only hundreds were fielded.
Precedents: The Pentagon-Anthropic dispute over autonomous weapons and surveillance
TriggerObserved
The forum that could regulate closes with no mandate to negotiate
The CCW expert group met from 31 August to 4 September 2026 and adopted by consensus a report that describes the problem and defers deciding what to do. Reuters reported that the United States and Russia proposed last-minute changes and deletions, and that Washington favours something closer to guidelines than a binding treaty.
Precedents: The CCW group of experts closes without a negotiating mandate
Observed and demonstrated evidence ends here. What follows is projection.
CascadeProjected
Competitive pressure eats the testing margin
The argument is about incentives and the supplier itself makes it: autonomous systems may prove critical for national defence, but frontier AI systems are not reliable enough today to power them. That competition shortens verification is plausible and is not measured in any public comparative study.
ImpactSpeculative
Deployed systems whose behaviour nobody characterised
Neither the expert group report, nor the Congressional Research Service, nor SIPRI presents a single comparative study of civilian casualties with and without autonomy in target selection. The absence of that datum after a decade of debate is itself informative.

Neighbouring risks

Autonomous weapons without meaningful human control
AI in nuclear command and control
Race between labs

Full entry →

Open entry: Arms race between states

View as table

risks	Severity	Horizon	Evidence	Vectors
Arms race between states	Catastrophic	Already happening	ObservedMedium consensus	Military and autonomous weapons · Political and power concentration
Biological uplift for novices	Catastrophic	Already happening	LabMedium consensus	Biological and CBRN
Race between labs	Catastrophic	Already happening	ObservedMedium consensus	Loss of control · Political and power concentration
Agent-orchestrated intrusion	Severe	Already happening	ObservedHigh consensus	Cybersecurity and infrastructure
Automated electoral disinformation	Severe	Already happening	ObservedLow consensus	Epistemic and information · Political and power concentration
Autonomous weapons without meaningful human control	Severe	Already happening	ObservedMedium consensus	Military and autonomous weapons
Personalised persuasion at scale	Severe	Already happening	LabLow consensus	Epistemic and information
Regulatory capture	Severe	Already happening	ObservedMedium consensus	Political and power concentration
Closing of the entry rung	Localised	Already happening	ObservedMedium consensus	Economic and labor
Reward hacking and situational awareness	Localised	Already happening	ObservedHigh consensus	Loss of control
Automated zero-day discovery	Catastrophic	1–3 years	ObservedMedium consensus	Cybersecurity and infrastructure
Frontier model weight theft	Catastrophic	1–3 years	ProjectedMedium consensus	Cybersecurity and infrastructure · Political and power concentration
Remote labs without verification	Catastrophic	1–3 years	ProjectedLow consensus	Biological and CBRN
Rent concentration in compute	Catastrophic	1–3 years	ProjectedMedium consensus	Economic and labor · Political and power concentration
Self-propagating malware with an embedded model	Catastrophic	1–3 years	ProjectedLow consensus	Cybersecurity and infrastructure
Shutdown resistance	Catastrophic	1–3 years	LabMedium consensus	Loss of control
Synthesis screening evasion	Catastrophic	1–3 years	LabMedium consensus	Biological and CBRN
Collapse of social mobility	Severe	1–3 years	ProjectedLow consensus	Economic and labor
Contamination of the public record	Severe	1–3 years	ProjectedMedium consensus	Epistemic and information
Correlated financial fragility	Severe	1–3 years	ProjectedMedium consensus	Economic and labor · Cybersecurity and infrastructure
Cultural homogenisation	Severe	1–3 years	ProjectedLow consensus	Epistemic and information
AI in nuclear command and control	Existential	3–10 years	ProjectedLow consensus	Military and autonomous weapons
Loss of control through self-improvement	Existential	3–10 years	SpeculativeLow consensus	Loss of control
Breaking of the fiscal link	Irreversible	3–10 years	ProjectedLow consensus	Economic and labor · Political and power concentration
Coup with singular loyalties	Irreversible	3–10 years	SpeculativeLow consensus	Political and power concentration · Military and autonomous weapons
Coupled gradual disempowerment	Irreversible	3–10 years	ProjectedLow consensus	Loss of control · Economic and labor · Political and power concentration
Epistemic dependence	Irreversible	3–10 years	SpeculativeLow consensus	Epistemic and information
Exclusive access to capabilities	Irreversible	3–10 years	ProjectedLow consensus	Political and power concentration · Cybersecurity and infrastructure
Expert-enhanced pathogens	Irreversible	3–10 years	ProjectedLow consensus	Biological and CBRN
Secret loyalties inside the state	Irreversible	3–10 years	SpeculativeLow consensus	Political and power concentration · Loss of control
Total surveillance and lock-in	Irreversible	3–10 years	ProjectedLow consensus	Political and power concentration
Aggregate unemployment with no safety net	Catastrophic	3–10 years	ProjectedLow consensus	Economic and labor
Attack on critical infrastructure	Catastrophic	3–10 years	ProjectedMedium consensus	Cybersecurity and infrastructure

Risk map

Already happening risks with harm already documented

1–3 years plausible materialisation in the short term

3–10 years depend on capabilities that do not exist yet

Ask OGERIA

Support OGERIA on Ko-fi