Frontier AI Safety
4.4x a year
Training compute for notable AI models has grown 4.4 times a year since 2010, about a doubling every six months.
Policy is eating the world.
As seen on
U.S. House of Representatives
United Nations
Perrin Research works on existential risk. Most of the work is on frontier AI.
The evidence
Where things stand.
Biosecurity
36of 38 providers
Researchers ordered camouflaged DNA fragments from 38 commercial synthesis providers. 36 shipped fragments of the 1918 influenza virus.
Nature Communications, Jan 2026Ordered with law enforcement's knowledge
Automated Decisions
67%
Two thirds of Medicare Advantage prior-authorisation denials that were appealed in 2025 were overturned.
KFF, Aug 202640% to 93% by insurer
Automated Decisions
<1% appealed
HealthCare.gov insurers denied 19% of in-network claims in 2024. Consumers appealed fewer than 1% of the denials.
KFF, Mar 2026At least 262,982 of about 85m
Cyber and Critical Systems
$820m
Ransomware operators were paid more than $820 million on-chain in 2025, 8% less than in 2024.
Chainalysis, Feb 2026May approach $900m as more is attributed
Cyber and Critical Systems
3,611complaints
The FBI's Internet Crime Complaint Center logged 3,611 ransomware complaints in 2025, with $32m in reported losses.
FBI IC3 Annual Report, 2025The FBI calls the loss figure artificially low
The pace
What frontier systems can do, and what goes wrong.
Source: METR, Task-Completion Time Horizons of Frontier AI Models, Last updated 8 May 2026. Each point is the state-of-the-art model at its release; confidence intervals are wide.
View the data
| Length of task, in the time an expert takes (log scale) | |
|---|---|
| GPT-2 | 3 sec |
| davinci-002 | 8 sec |
| GPT-3.5 Turbo Instruct | 36 sec |
| GPT-4 | 4 min |
| GPT-4o | 7 min |
| Claude 3.5 Sonnet | 11 min |
| o1-preview | 20 min |
| o1 | 39 min |
| Claude 3.7 Sonnet | 1 hr |
| o3 | 2 hr |
| GPT-5 | 3.4 hr |
| Claude Opus 4.5 | 4.9 hr |
| GPT-5.2 | 5.9 hr |
| Claude Opus 4.6 | 12 hr |
| Claude Mythos Preview | 17 hr |
Source: AI Incident Database via Stanford HAI, AI Index 2026, fig. 3.2.1, April 2026 (9th edition). Redrawn from the AI Index public data, CC BY-ND 4.0.
View the data
| Incidents per year | |
|---|---|
| 2014 | 13 |
| 2015 | 24 |
| 2016 | 41 |
| 2017 | 50 |
| 2018 | 45 |
| 2019 | 43 |
| 2020 | 90 |
| 2021 | 77 |
| 2022 | 104 |
| 2023 | 168 |
| 2024 | 281 |
| 2025 | 362 |
The record
What we have published.
From the BoardOctober 1, 2026
Strategy, leadership and governance for FY 2027The Board of Directors of Perrin Institution, Inc. has issued its official notice for FY 2027: four priorities for the year, the election of a Chairman, and the governance provisions that apply to the organization.
BriefingRevised Aug 2026
Explainable AI and the limits of regulationBriefingRevised Aug 2026
Why federal privacy legislation has not passed<1%of denied claims appealed
Reference buildIn development
Overturn: an agent that appeals an automated denialLegislationSigned Dec 19, 2025
The RAISE Act, signed into law in New York15governance documents
Governance2026
Fifteen documents on how the institution is run and funded
Archive2023 to 2025
383 publications from the Perrin & Fidutam record, collected in a 1,549-page compendiumFrom the Board
The Institution exists to provide stakeholders and legislators with the best frameworks possible to mitigate existential risk within the modern superintelligence exponential.
Briefing
Explainability is not a property of a model. It is a relationship between a model and the person who has to live with its decision.
Briefing
The United States does not lack a federal privacy law by accident. It has repeatedly chosen sectoral rules, and the cost of that choice now falls on data the sectors never anticipated.