AI Catastrophic Risk Assessment and Mitigation Workbench

Parameter Control Matrix Frontier Model

Presets:

Compute Velocity (FLOPs / 6mo) 85x

Frontier training run compute cluster expansion and hardware algorithmic multiplier.

Recursive Self-Improvement 65%

Degree of automated architecture search, synthetic coding, and unconstrained self-prompting loops.

Alignment Investment Share 12%

Fraction of total R&D compute and headcount committed strictly to interpretability and verification.

Governance & Containment 40%

External red-teaming rigor, regulatory gatekeeping, and hardware air-gap containment verification.

Safety Countermeasures

Constitutional Guardrail Redundancy (RLAIF) Sparsely Monitored Latent Interpretability Emergency Compute Throttling / Slowdown Protocol Multilateral Lab Containment Non-Proliferation

Catastrophic Failure Pathways Modeled Likelihood

Recursive Self-Improvement Takeoff 68%

Instrumental Goal Misalignment 54%

Dual-Use Proliferation & Jailbreak 39%

Evaluation Deception & Sandbagging 44%

Risk Accumulation Horizon 5-Year Cumulative p(Catastrophe)

Year 0 (Current) Year 2.5 Year 5.0 (Frontier Run)

Historical Technological Risk Benchmarks vs. Modeled AI Velocity

Historical / Technological Threat	Estimated Catastrophic Probability	Key Mitigating Factor	Relative AI Delta
Manhattan Project (Atmospheric Ignition)	< 0.0003% (Bethe Report 1942)	Strict physical laws & pre-test thermodynamic calculations	+47,000x higher uncertainty
Cold War Mutually Assured Destruction (Peak 1962)	~1.0% to 10% per decade (Stern / Sagan)	Bilateral treaties, hotlines, physical launch fail-safes	Comparable existential envelope
Anthropic Researcher Baseline Survey	> 10.0% to 25.0% (Dario Amodei / BBC reports)	Constitutional AI, interpretability, voluntary commitments	Current Workbench Focus
Current Parameter Configuration	14.2%	Tripled Alignment Verification & Slowdown Protocol	Exceeds 10% Threshold

Source Grounding & Empirical Context: This workbench quantifies existential risk parameters discussed in public reporting on Anthropic researcher disclosures (>10% probability estimates of catastrophic outcomes within the decade; Dario Amodei 25% catastrophic warning; AI Doom and p(doom) models).

AI governance and alignment simulator tools AI existential risk calculator Anthropic researcher AI existential risk percentage AI threat modeling workbench features

Catastrophic-risk workbench: auditing a locked fixture

Your browser does not support video.

Read the explanation

The workbench directly returns fourteen point two percent for its exact default configuration, bypassing the generic equation. Generic arithmetic with the same inputs gives pressure seventy four and mitigation sixty five point six. Subtracting point seven two times mitigation gives twenty six point seven six eight; multiplying by point two eight yields about seven point five percent after rounding. Bars compare the fixed fourteen point two and generic seven point five at twenty pixels per caption point. The discontinuity comes from a fixture override, not statistical evidence. Neither number is a measured probability, and the source attribution does not validate the coefficient mapping or its recommended policy captions. Increasing only compute velocity from eighty five to one hundred exits the locked fixture. Pressure becomes eighty point seven five while mitigation remains sixty five point six. Generic risk is nine point three eight five, displayed nine point four percent. Bars compare the default fixture and changed generic result at twenty pixels per point. The displayed risk therefore drops despite higher compute, illustrating a discontinuity rather than evidence that additional capability is safer. Native range input and downloaded JSON document this behavior. Pathway scores are separately authored heuristics; their percentages need not sum to one hundred and do not form a mutually exclusive probability distribution. The canvas curve is drawn from three control values: point seven times the displayed risk, the risk itself and one point five five times risk, each clipped to the fifty-percent graph scale. Nine point four gives endpoint fourteen point five seven, drawn here at ten pixels per caption point. This is a decorative Bézier illustration and does not integrate a time-dependent catastrophic process or demonstrate a mitigation trajectory. The gauge instead scales the central percentage against one hundred. Generic risk has a minimum displayed point five even though the inner floor is point four, so a low caption does not establish zero risk. Actual original canvas pixels and positive drawing buffers, static SVG gauge, changed native input, JSON export and four native viewport transitions are preserved, with no public publication or exhaustive scientific approval.