828dda016c
- Source: inbox/archive/2025-07-00-fli-ai-safety-index-summer-2025.md - Domain: ai-alignment - Extracted by: headless extraction cron (worker 6) Pentagon-Agent: Theseus <HEADLESS>
2.9 KiB
| type | domain | description | confidence | source | created | secondary_domains | |
|---|---|---|---|---|---|---|---|
| claim | ai-alignment | Only Anthropic, OpenAI, and DeepMind conduct substantive testing for dangerous capabilities among frontier AI developers, per FLI Summer 2025 index | likely | Future of Life Institute, AI Safety Index Summer 2025, July 2025 | 2026-03-11 |
|
Frontier AI dangerous capability testing is limited to three companies
Future of Life Institute's Summer 2025 evaluation found that only three frontier AI companies—Anthropic, OpenAI, and Google DeepMind—conduct substantive testing for dangerous capabilities such as bioterrorism potential and cyberattack vectors. The remaining four assessed companies (x.AI, Meta, Zhipu AI, DeepSeek) either do not test for these risks or do not disclose such testing.
This concentration of risk assessment creates a bifurcated landscape where the majority of frontier AI development proceeds without systematic evaluation of catastrophic misuse potential. The gap is particularly concerning given that all companies are developing increasingly capable systems, but only a minority are actively probing for dangerous emergent capabilities.
The FLI index evaluated companies across six dimensions, with "Risk Assessment" specifically measuring dangerous capability testing. The three companies conducting such testing still received overall grades of C+ to C-, indicating that even substantive risk assessment does not guarantee comprehensive safety practices.
Why this matters: Dangerous capability testing remains a voluntary practice adopted by safety-conscious labs rather than an industry norm, leaving significant capability development unmonitored for catastrophic risk vectors. This pattern reinforces the structural race-to-the-bottom dynamic: companies that invest in expensive risk assessment (bioterrorism, cyberattack modeling) gain no competitive advantage and may face capability delays, while competitors skip these costs entirely.
Evidence
From FLI's assessment:
- 3 of 7 frontier AI companies (Anthropic, OpenAI, DeepMind) conduct substantive dangerous capability testing
- 4 of 7 companies (x.AI, Meta, Zhipu AI, DeepSeek) do not conduct or disclose such testing
- Testing focuses on bioterrorism potential and cyberattack capabilities
- Even companies conducting testing scored C+ or below overall
Relevant Notes:
- AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur which makes bioterrorism the most proximate AI-enabled existential risk
- voluntary safety pledges cannot survive competitive pressure because unilateral commitments are structurally punished when competitors advance without equivalent constraints
- the alignment tax creates a structural race to the bottom because safety training costs capability and rational competitors skip it
Topics: