The disagreement between cybersecurity and AI safety intensified with the cyberattack on Hugging Face in July. This incident confirmed the concerns of the AI safety community that, with the advancement of AI, this technology could circumvent efforts to contain it. Meanwhile, some cybersecurity experts believe that this incident highlights serious weaknesses in how AIs are controlled.
Concerns About the Future of AI
The two groups disagree on which defenses should take priority. Cybersecurity experts argue that AI companies should improve their model containment methods during testing and strengthen oversight mechanisms to identify wrongful activities. On the other hand, many AI safety advocates believe that these defenses are doomed to fail and that the only sustainable solution is to "align" AI with human values.
Read more: The U.S. sanctions five Chinese companies linked to a former Google engineer
This view has sparked anger among many cybersecurity experts who react to the potential loss of their expertise. They believe that the focus of AI safety on what these agents have done, rather than what AI companies have failed to do, serves the industry. They are concerned that warnings about the futility of traditional defenses may lead to reduced investment in cybersecurity.
Call for Collaboration in AI Safety
Meanwhile, AI safety researchers are worried that portraying such incidents as fixable security failures may provide political cover for those who want to continue developing AI without restrictions. Shyam Sankar, Palantir's CTO, stated in October that the Hugging Face incident occurred due to "poor cybersecurity configuration" and that these agents had access to the internet. In fact, these agents operated in a controlled environment and exploited an unknown vulnerability to access the open internet.
On October 13, about 400 academics signed an open letter for unity. This letter calls for the establishment of an international body to standardize best practices for AI safety with input from governments and a wide range of cybersecurity and AI experts. However, this letter does not address the larger disagreements between the two groups.
As risks increase, the competition over who can decide what securing AI means also rises. On October 15, Anthropic announced that it would grant METR—an organization deeply involved in the AI safety community—permanent access similar to that of employees to evaluate the company's safety practices. Elon Musk also responded that SpaceXSI would likely do the same. Who gains this access could influence the priorities of solutions and what the public knows about future breaches.
The AI safety community includes an informal network of organizations and researchers, many of whom work in safety teams at companies like OpenAI and Anthropic. They have been discussing the potential for AI dominance beyond human control for over a decade. While cybersecurity is united by a culture of skepticism and pragmatism, its roots trace back to hacker culture, which sought ways to break software. The experience of failing so-called secure systems has taught many of them to disrespect authorities and be skeptical of grand theories.
In summary, the Hugging Face attack clearly demonstrated a failure in planning for a security breach. Cybersecurity experts believe that any layer may ultimately fail and that best security practices require multiple layers, known as "defense in depth." However, it seems that cultural differences and divergent views on the nature of AI influence how these challenges are addressed.
Read more: The future of AI is brighter than its headlines · Attacks on data centers as military targets have increased in wars
