An analysis by the UK’s AI Safety Institute finds that the most capable open-weight artificial intelligence models are narrowing the gap with the best closed models in cyber-related capabilities. According to the institute, the top open models lag the leading closed models by about four to seven months.
Benefits and risks
Open-weight models offer clear advantages: they can be modified to suit users' needs and facilitate easier collaboration because weights and configurations are accessible. At the same time, those same characteristics make it possible to remove guardrails, potentially making powerful cyber and biological capabilities more accessible to bad actors.
Examples and concern
The New York Times has reported that Boko Haram militants in Nigeria are using (mostly closed) AI to plan attacks and gain tactical advantages. Such incidents illustrate that AI-enabled capabilities are already being exploited for militant or criminal purposes.
Anthropic’s CEO recently wrote that open models “potentially present a higher risk,” while also suggesting that all powerful models — not only open ones — should undergo safety testing.
Implications
The analysis implies that as open models advance quickly, regulation, safety testing, and international cooperation become increasingly important. Democratization of access to these technologies brings research and innovation benefits, but also raises the likelihood that malicious actors could exploit advanced capabilities. The four- to seven-month gap provides policymakers and the research community with a concrete time horizon for adaptation before open models potentially match the current capabilities of closed systems.
— Source: analysis by the UK’s AI Safety Institute; article by Tom Chivers
