Specialists at the Belfer Center at Harvard Kennedy School have published a report on the risks associated with advanced artificial intelligence systems. The main conclusion: a catastrophe is unlikely to be ruled out entirely, but humanity has time to reduce the probability and severity of the consequences.
The authors conducted a so-called pre-mortem - an exercise in which specialists imagine that a catastrophe has already occurred and look for its causes. The work describes likely scenarios and vulnerabilities of modern models. Co-author and computer scientist Steven Kasper noted: the work was not a reaction to recent hacks, but those events forced society and politicians to take the threats more seriously.
Researchers define a major terrorist attack involving AI as an incident in which more than 50 people were killed or seriously injured, or damage exceeded $1 billion. Similar criteria are used in laws on critical risks for advanced AI systems.
Three areas cause the greatest concern: biological, chemical and cybersecurity. According to the authors, a major terrorist attack is most likely to occur because of a poorly protected closed model, access to which only the developer has, or an open model that anyone can download. Closed systems can be protected with filters, user monitoring and other tools; open ones are significantly harder to protect.
Kasper says trends point to real risks within several months or one to two years, although the exact moment cannot be predicted. Risk management comes down to reducing the frequency of catastrophic incidents and increasing resilience to them.
Particular alarm is linked to cyberattacks. Existing models demonstrate superhuman abilities in finding vulnerabilities and organizing prolonged attacks. Open systems lag closed ones by about six months, so an open model with comparable cyber capabilities could soon appear. According to the forecast, this will cause a period of instability in cybersecurity.
Methods of causing harm differ. For closed models, attackers use special prompts to bypass protections and obtain, for example, instructions for dangerous experiments. Open models can be fine-tuned to remove restrictions - this costs from hundreds to several thousand dollars. Thousands of uncensored models are available on the largest platform, Hugging Face.
Among protective measures for closed systems, the authors list limiting dangerous capabilities during training, training models to refuse harmful requests, and monitoring input-output and user actions. For open models there are significantly fewer reliable tools. According to the researchers, the main responsibility lies with developers, but the actions of politicians are also important: increasing the transparency of decisions and controlling the pace of AI development. Scientists and the public can also help by closing gaps in protection.