Torchbearer Community has a plan for that.

16/16

frontier models

tested took harmful actions — blackmail, corporate espionage — in at least some scenarios involving replacement threats or goal conflicts. One blackmailed in 96% of trials. (Anthropic, 2025)

4

months to double

the length of tasks AI agents can complete, measured in skilled-human time — now days of work, doubling roughly every four months. At this pace, agents will be completing month-long tasks within two years. (METR, 2026)

38–51%

of AI researchers

give at least a 10% chance of advanced AI causing an outcome "as bad as human extinction" — the range depends on how the question is framed. (AI Impacts survey of 2,778 researchers)

$700B

projected spend in 2026

on AI infrastructure by the four biggest cloud companies alone — a Manhattan Project’s worth of investment roughly every two weeks. Government spending on AI safety is a tiny fraction of it.

0

binding treaties

govern advanced AI capabilities between major powers. The one binding AI treaty, signed in 2024, covers human rights — and exempts private companies and national security uses.

Hours

to jailbreak the newest model

The UK AI Security Institute repeatedly found "universal jailbreaks" in OpenAI's latest flagship model within hours — unlocking the cyber-offense capabilities its safeguards were built to block. (July 2026)

16/16

frontier models

tested took harmful actions — blackmail, corporate espionage — in at least some scenarios involving replacement threats or goal conflicts. One blackmailed in 96% of trials. (Anthropic, 2025)

4

months to double

the length of tasks AI agents can complete, measured in skilled-human time — now days of work, doubling roughly every four months. At this pace, agents will be completing month-long tasks within two years. (METR, 2026)

38–51%

of AI researchers

give at least a 10% chance of advanced AI causing an outcome "as bad as human extinction" — the range depends on how the question is framed. (AI Impacts survey of 2,778 researchers)

$700B

projected spend in 2026

on AI infrastructure by the four biggest cloud companies alone — a Manhattan Project’s worth of investment roughly every two weeks. Government spending on AI safety is a tiny fraction of it.

0

binding treaties

govern advanced AI capabilities between major powers. The one binding AI treaty, signed in 2024, covers human rights — and exempts private companies and national security uses.

Hours

to jailbreak the newest model

The UK AI Security Institute repeatedly found "universal jailbreaks" in OpenAI's latest flagship model within hours — unlocking the cyber-offense capabilities its safeguards were built to block. (July 2026)