ATxSummit Leaders Warn AI Safety Cannot Wait for Catastrophe

Stuart Russell has spent decades thinking about machine failure modes. The UC Berkeley computer scientist, whose textbook Artificial Intelligence: A Modern Approach shaped a generation of researchers, told the ATxSummit conference in Singapore this week that the industry cannot afford to wait for catastrophe before acting. His framing was stark: a Chernobyl-scale disaster would not merely trigger regulation. It would trigger a societal shutdown.

"Without safety, there are no benefits," Russell told the panel. The trillions of dollars currently flowing into AI development would be stranded if public trust collapsed in the wake of a major incident. Computer Weekly covered the Singapore gathering, where panelists described AI governance as entering an urgent new phase.

The central question has already shifted. Whether artificial intelligence needs oversight is no longer seriously contested. The live debate is whether governments and industry can build accountability structures fast enough to keep pace with systems that are growing more capable, harder to audit, and more deeply embedded in daily life.

The speed problem

Karan Bhatia, Google's global head of government affairs, put the structural challenge plainly. Traditional regulatory timelines bear no relationship to how quickly frontier models are being developed and deployed. He called for a permanent, structured dialogue between regulators and industry: routine sharing of threat intelligence, continuous iteration on policy options, not episodic crisis summits.

That urgency takes on concrete meaning given recent events at the capability frontier. PBS NewsHour reported last month that Anthropic began limited testing of a model called Mythos, which the company itself described as potentially causing widespread disruption if released broadly. The model is reportedly capable of sustained, multi-step security research at a pace that outstrips what a skilled human professional could accomplish in a full working day. More than 40 companies, including some competitors, were granted access to probe for vulnerabilities.

Even that controlled rollout raised concerns. Capabilities that can identify exploitable software gaps at scale do not require bad actors to cause harm. They require only misconfiguration, or a gap between what a system can do and what its developers anticipated.

Governance by design

Elham Tabassi, director of the AI and Emerging Tech Initiative at the Brookings Institution, offered a structural answer to reactive regulation. Governance must be embedded in the development process from the start, she argued, not applied as a post-release audit. Trustworthiness by design, not by remediation.

That principle reflects a broader shift in how practitioners approach an artificial intelligence review of safety-critical systems. For years, safety work was treated as secondary to performance benchmarks. The ATxSummit panel suggests that framing is losing credibility among those close enough to the technology to understand the asymmetry of risk: the upside of moving fast is incremental, the downside of a major failure potentially permanent.

Tracking model releases across the industry, LLM Stats and Price Per Token both show frontier systems shipping at a pace that governance frameworks simply cannot match. New capabilities are arriving faster than the ink dries on proposed rules.

Russell's Chernobyl analogy carries specific historical weight. The 1986 disaster did not just force safety reforms on the nuclear industry. It effectively ended civilian nuclear expansion across much of the Western world for a generation. An equivalent AI incident would carry that kind of irreversibility, and the damage would extend beyond the companies directly involved to every researcher building toward applications in medicine, scientific computing, or public infrastructure.

What the ATxSummit discussion leaves open is the mechanism. Bhatia's proposal for continuous regulator-industry contact assumes regulators have technical staff capable of engaging substantively on an ongoing basis. Most do not. Tabassi's governance-by-design principle is sound but depends on developers accepting meaningful oversight before they fully understand what their systems will eventually become capable of. Both proposals require institutional trust that does not yet exist at the necessary scale.

Building that trust before a serious incident forces the issue is the practical challenge the field now faces. Whether the current generation of leaders has the political will to make it happen is the harder question.

Frequently asked questions

What did panelists mean by a "Chernobyl moment" for AI?
They warned that waiting for a catastrophic AI incident before building governance frameworks would mirror the post-Chernobyl pattern in nuclear energy, where belated reform came too late to preserve public confidence in the technology for decades.

Who is Stuart Russell and why does his view carry weight?
Russell is a distinguished computer science professor at UC Berkeley and co-author of Artificial Intelligence: A Modern Approach, the field's foundational textbook. His long-standing focus on AI safety and existential risk gives his warnings particular credibility among researchers.

What is Anthropic's Mythos model?
Mythos is a highly capable model withheld from public release due to concerns about misuse. Anthropic is sharing access with more than 40 firms to test for security vulnerabilities, specifically around its ability to identify exploitable software flaws at a scale that surpasses skilled human researchers.

What does "governance by design" mean for AI developers?
It means integrating safety and accountability mechanisms into the development process from the start, rather than auditing systems after they are already deployed. Proponents argue this approach is fundamentally more effective than reactive regulation applied after capabilities are already in the wild.