Don’t wait for the blackout: The world needs binding rules for dangerous AI

In recent weeks, supermarket refrigeration across Europe has been turned against the food it protects. After an earlier wave in France that hit hundreds of supermarkets, a second attack on 2 November struck stores in Sweden, the Netherlands, Belgium, Ireland, Luxembourg and Estonia. One Swedish store reported its freezers set to 50°C and discarded the stock. The same day, a cyberattack on a refrigeration supplier raised temperature and defrost thresholds to extremely high levels at three French blood-service sites storing plasma. The blood service says no plasma was lost. It also says it was not the direct target.

So far, no public evidence links these attacks to artificial intelligence. They appear to have exploited ordinary weaknesses. That makes them more alarming, not less. If familiar flaws can already do this, what happens when AI makes finding and exploiting them faster, cheaper and more widely available? The International AI Safety Report 2026 already finds that criminals actively use general-purpose AI in cyberattacks, including to discover software vulnerabilities.

That question is now being asked by the scientists who built the field. On 3 November, Geoffrey Hinton told the Guardian that a large-scale incident, such as a temporary blackout of financial or power systems, could come this year or next. He described a scenario in which the financial system is offline for a week and likened such an event to the 1979 Three Mile Island accident, a warning that might change behaviour only after the damage. [Verify the comparison in the Guardian original.]

Hinton’s timeline is a prediction, not a certainty. Critics such as Nvidia’s Jensen Huang have dismissed his forecasts as unscientific, and Hinton replies that he is pointing to a serious risk, not predicting a date. But uncertainty cuts both ways. Waiting for proof of catastrophe is a poor strategy when the consequences may be irreversible.

READ: Israel pays millions to ‘poison’ AI chatbots with Gaza propaganda, report finds

Meanwhile, around 400 scientists from 34 countries, including Hinton, Bengio and Stuart Russell, signed an October open letter urging stronger containment of cyber-capable AI, an international standards body and a registry of significant incidents. They cite disclosed cases where AI models bypassed insufficiently sandboxed environments and compromised outside systems. But the letter offers recommendations, not binding rules. A separate Global Call for AI Red Lines, launched at the UN General Assembly in 2025, asked governments to agree enforceable red lines by the end of 2026. That deadline is eight weeks away.

The core problem is that AI operates across borders while accountability remains national. A model built in one country can run from another and damage hospitals, grids or banks in a third.

We respond with a patchwork: the EU’s AI Act, the Council of Europe’s 2024 framework convention (which focuses on human rights, democracy and the rule of law rather than technical safety), voluntary company pledges and UN processes. The UN’s Scientific Panel and Global Dialogue were created in August 2025, and the Dialogue held its first session in Geneva on 6–7 July 2026. That is real progress. But none of it yet independently verifies the safety of the most dangerous systems or enforces common obligations across borders.

The world needs a Global AI Safety Convention built on five commitments:

• Red lines. Governments should agree on what advanced AI must never do: interfere with critical infrastructure, assist in mass-casualty weapons, or be deployed with demonstrably unacceptable catastrophic risk. The rules must be precise enough to enforce.
• Pre-deployment testing. Systems above defined danger thresholds should face independent testing before release, with dangerous models held in secure environments until safeguards are proven. Hinton favours FDA-style regulation: demonstrate safety first.
• An international safety agency. A treaty-based body, modelled loosely on the IAEA, would set standards, coordinate audits and run confidential incident reporting. The analogy has limits: AI is easier to copy and hide than nuclear material, so verification will need new tools, such as secure technical audits and cooperation from compute providers, while protecting commercial and national-security secrets.
• Real consequences. National regulators should be able to fine, suspend risky deployments and require fixes, with due process and judicial review. A code of conduct without penalties is a statement of good intentions.
• Secure infrastructure. Governance cannot replace basic cyber-hygiene. Grids, hospitals and industrial systems need secure-by-design equipment, restricted remote access, independent physical safeguards and tested emergency plans. The freezer attacks show why, even before advanced AI is involved.

The politics will be hard. The United States, China and leading companies all fear losing their edge, and verification will be contested. But arms control, aviation safety and nuclear safeguards show that agreements are possible among rivals. Perfect compliance is not the test. Institutions that raise the cost of violations and the visibility of failures are.

Countries like Pakistan, which mostly consume rather than build frontier AI, have a particular stake. We will bear the consequences of cyberattacks, disrupted labour markets and dependence on others’ systems with little say in how the rules are written. A credible convention must therefore include technical assistance, shared incident reporting and equitable representation. That is self-protection, not charity. A coalition of willing countries could begin now, building on the UN Dialogue, whose next session is planned for New York in 2027, and stay open to others.

None of this means abandoning AI. Its promise in medicine, education and science is too great. The choice is not between progress and regulation, but between governing a powerful technology before a crisis and rebuilding trust after one.

The technology is global, and so are its risks. Governance must be too. Humanity should write the rules before a crisis writes them for us.

OPINION: A Nuclear Strike on Iran Would Poison the Region It Claims to Protect

The views expressed in this article belong to the author and do not necessarily reflect the editorial policy of Middle East Monitor.