Debates about AI safety often swing between predictions of extreme catastrophe and concerns about harms already taking shape. A panel of experts interviewed by WIRED described a more specific set of risks: cyberattacks, biological threats and the adoption of error-prone AI in nuclear operations. Their views also point to a harder question: who gets to decide what counts as safe enough?
Concrete risks are easier to examine
Several experts in the panel identified cybersecurity and biological attacks as serious concerns. AI may help defenders improve their systems, one expert said, but infrastructure may not adopt new defenses quickly enough to keep pace with attackers. The same expert described the technology as potentially tilting toward attackers in the near term.
That concern does not mean every improvement in AI will inevitably cause harm. One panelist argued that safety discussions should identify the specific danger first, because different threats call for different defenses. In cybersecurity, the expert pointed to bug bounties as an established way to reward people who disclose attack methods or vulnerabilities.
For biological security, the panelist raised screening orders for DNA fragments as a possible safeguard against misuse. The proposal illustrates the broader point: a useful safety response depends on what people are trying to prevent. General fears about increasingly capable systems do not, on their own, specify what protections are needed.
Nuclear operations raise reliability concerns
One expert focused on the prospect of AI systems entering nuclear workflows and operations. The concern was not framed as a prediction that catastrophe is certain. It was about integrating systems described as faulty and unreliable into settings where failures could have catastrophic consequences.
The expert also challenged the way some people compare AI with nuclear technology while rejecting established risk assessment methods as inadequate. The panelist described risk assessment as a field shaped in part by the Manhattan Project, with standards intended to limit harm to people, the environment, and financial and national security assets.
At the center of that approach is a public question: how safe is safe enough? The expert argued that safety claims should be verifiable and falsifiable, and linked assessment to estimating how many people might die. From this perspective, claims about human extinction lack the grounding needed for conventional safety engineering when no supporting numbers are offered.
Disagreement over catastrophic scenarios
The panel was not dismissive of AI risks. Its members expressed concerns about cyber and biological attacks, and about AI in defense. But they did not all agree on how likely the most sweeping scenarios are or how quickly change will unfold.
One expert called extinction a particularly overblown risk, arguing that safety work should focus on specific, testable harms. Another described disagreement with the coauthors of AI 2027, who expect an extraordinary amount of change, while predicting a dramatic but less extreme impact. These differences show why the debate needs to distinguish between the scale of possible outcomes and the evidence for particular claims.
That distinction also shapes the argument over AI progress itself. The cybersecurity expert rejected the idea that smarter AI automatically produces downstream harm. Instead, the expert argued, risks should be named individually so that defenses can be designed to address them.
Who sets the rules?
Most members of the panel were skeptical that a pause in AI development would last. They pointed to the close ties between major AI companies and capitalism as a reason development could resume or continue. The interviews also repeatedly raised democracy and governance, including concerns that AI’s current direction could be tyrannical.
Lawmakers have a role, but some experts expect governments to move too slowly. The article reports one technologist’s skepticism that US courts can serve as arbiters of global AI safety. Yoshua Bengio, meanwhile, urged the public to pressure politicians, saying, “That’s your job too.”
For the cybersecurity expert, there was broad agreement on some responses even amid disagreement about AI’s future: companies should be more transparent, face liability when things go wrong, and protect whistle-blowers inside AI companies. Those proposals make safety a question of public accountability as well as technical reliability.
The panel’s answers do not settle whether AI poses an existential threat. They do offer a way to make the debate more concrete: identify the harms, assess how systems could contribute to them, and decide through public institutions what level of risk is acceptable. As the article’s theologian argues, concerns about losing control of essential institutions deserve attention, while questions about consciousness, creativity and human connection remain open. The future of AI safety, in this view, cannot be left entirely to corporations.