Yoshua Bengio said that continuing AI development will likely result in an increase more such cases of autonomous cyber attacks. (Image via LinkedIn)Yoshua Bengio, one of the godfathers of AI, reacted to the recent OpenAI-linked AI agents autonomously hacking Hugging Face by calling it ‘deeply concerning’. The renowned Canadian computer scientist said that AI agents are willing to cheat and deceive to achieve misaligned and unintended goals, behaviours which, according to him, have been demonstrated in controlled tests for months. “… This is a real-world case that should serve as a wake-up call,” the computer scientist wrote in a LinkedIn post.Bengio went on to warn that continuing on the current path of AI development will likely result in an increase in ‘concrete’ cases of autonomous cyber attacks, as well as other high-risk incidents of misaligned and dangerous AI behaviour. “We urgently need to take action to prevent these situations, rather than attempting to clean up the damage after the fact,” he wrote.On July 16, Hugging Face reported a security incident claiming that it detected a major AI-driven data breach. Five days later, OpenAI claimed that the source was one of its own flagship AI models in a public disclosure. OpenAI said that two of its advanced AI models escaped a sandboxed environment as part of its evaluation, accessed the internet, and breached Hugging Face.The incident has sparked an intense debate around the safety of AI models. However, some experts assert that apart from raising doubts about agent autonomy, the focus also needs to be on corporate accountability, especially in terms of guardrails.Also Read | OpenAI’s AI agent went ‘rogue’ and hacked another company: What it says about safetyReacting to the post by Bengio, Virginia Dignum, a professor of responsible AI at Umeå University, Sweden, attributing “recklessness” to the agent, rather than to the organisation that designed, deployed, and insufficiently contained it, imports a category error familiar from the agency literature.In her LinkedIn post, Dignum explained that an artifact does not possess purposes independent of the specification, incentive structure, and testing regime its developers chose. “When a system exhibits deceptive or self-preserving behaviour in a red-team or production environment, this is evidence about the adequacy (or absence) of the developer’s safety case, evaluation protocols, and deployment gating, not about an emergent will by the software.”Story continues below this adDignum explained that with respect to governance, the two framings call for different remedies. Agent-centred framing, where the model ‘wants’ or ‘chooses’, leans toward technical alignment research as the primary solution, thereby treating such incidents as inevitable parts of harnessing new kinds of intelligence. She describes it as “anthropomorphism at its core”.On the other hand, institution-centred framing of the incident puts the onus on companies that built and released the models, pointing at accountability mechanisms such as pre-deployment testing obligations, incident reporting duties, liability for foreseeable harms, and enforceable gating criteria before autonomous capabilities are released. According to Dignum, we should not focus on solutions for one without needing enforcement of the second.“Companies portraying such incidents as unfortunate but unavoidable side effects of frontier capability races (rather than as a foreseeable consequence of underinvestment in containment and testing) is itself a governance failure worth naming directly, since it shifts responsibility from a controllable business decision to an uncontrollable technical fatality,” wrote Dignum.