Three AI risks that do not require superintelligence.
This is a special edition of AI, Built Human. On Tuesday, October 6, I am speaking at an open meeting on AI in politics hosted by Radikale Venstre. Preparing it produced more material than fits into one evening, and the topic cannot wait for the next regular edition. These are my notes in text. There is also a 23-minute podcast version on Substack.
In 2017, I was working as a consultant at the headquarters of a large Danish shipping company when ransomware hit.
I watched screens go black, one after another, down the rows. I managed to disconnect my own computer from the network.
That image stayed with me while preparing the talk. The invitation mentioned the end of humanity. I ended up somewhere more down to earth: the risks that should keep politicians awake do not require superintelligence. One of them is already hitting Danish water utilities.
Three near-term AI risks
None requires superintelligence.
| Risk | Status | Political action |
|---|---|---|
| Bioweapons: The companies themselves are raising the alarm | Several AI developers introduced extra safeguards in 2025. | Who may release models, who tests them, who controls synthetic DNA? |
| Division: Algorithms have already shown what they can do | In experiments, AI content can shift opinions as effectively as human-written content. | Enforce the DSA and Article 50 of the AI Act. |
| Cyberattacks: Already happening in Denmark | Energy and water have been hit, and the threat has been raised to high. | Defence, preparedness and NIS2. |
How urgent? Bioweapons and division require something we have not yet seen at scale. Cyber: it is underway.
Cyber is the most urgent of the three.
1. The long-term scenario is uncertain. The near-term risks are documented
AGI and superintelligence are used very loosely. AGI is AI with human-like breadth, able to learn and solve many different tasks at least at human level. Superintelligence is AI that far surpasses the best humans in almost every area. Both terms describe capabilities. They say nothing about consciousness, and nothing about safety.
The international AI safety report from February 2026 describes early signs in tests of models bypassing oversight to achieve a goal. It also concludes that systems do not yet have sufficient capabilities for an actual loss of control, and that researchers strongly disagree on its probability.
In his 2024 Nobel banquet speech, Geoffrey Hinton pointed to two shorter paths: AI has already created divisive echo chambers, and in the near future it could be used to create new viruses (freely translated). I add a third.
2. Bioweapons: The companies themselves are raising the alarm
According to the safety report, several AI developers introduced extra safeguards in 2025 because they could not rule out their models helping a novice with biological weapons.
Open-weight models make the problem larger. They cannot be recalled, safeguards are easier to remove, and they can run completely unsupervised. Materials, equipment and practical skills remain a barrier. The question is for how long.
The chain has many links, several of them political: who may release which models, who tests them, and who controls orders for synthetic DNA?
3. Division: Algorithms have already shown what they can do
Pew Research found in 2022 that 72 percent of American Republicans and 63 percent of Democrats considered the other side more immoral than other Americans.
That happened with algorithms that only choose what we see. With AI, content is also written, conversations conducted and everything tailored to the individual. The safety report shows that AI-generated content can be as effective as human-written content at shifting opinions in experiments.
Some of the tools already exist. The EU Digital Services Act requires the largest platforms to mitigate risks to elections and public debate. Article 50 of the AI Act on labelling and transparency applies from August 2026. The political task is to enforce them.
4. Cyberattacks: Already happening in Denmark
In May 2023, 22 Danish energy companies were attacked simultaneously. In late 2024, pro-Russian hackers manipulated the controls at a Danish water utility, and the Danish Defence Intelligence Service has since linked the group to the Russian state. On September 24, 2026, the Danish Agency for Societal Security raised the threat of destructive cyberattacks from medium to high. Energy and water are specifically mentioned.
AI is already involved. In November 2025, Anthropic reported that a Chinese state-backed group had used its AI for 80 to 90 percent of a hacking campaign against around 30 targets. The campaign was detected and stopped because it ran on Anthropic's servers.
The threat is already here
Documented figures from Denmark and the attacking side.
| Number | Meaning | Source |
|---|---|---|
| 22 | Danish energy companies attacked simultaneously in May 2023 | SektorCERT, November 2023 |
| 80–90 % | of a state-backed hacking campaign carried out with AI | Anthropic, November 2025 |
| approx. 30 | targets in the same campaign, which was detected and stopped | Anthropic, November 2025 |
The threat of destructive cyberattacks has been raised to high. Danish Agency for Societal Security, September 24, 2026. Energy and water are specifically mentioned.
Model 1. Cyber is the risk already underway.
With an open-weight model on their own machines, nobody sees anything. Since summer 2025, Chinese open-weight models have overtaken American ones in download numbers. A hacking unit can download one, remove the safeguards and run it locally. There is then no provider who can pull the plug.
That is why I call cyber the most urgent of the three. The first two require something we have not yet seen at scale. The third is underway.
The first two require something we have not yet seen at scale. The third is underway.
5. We cannot regulate Chinese models. We can defend ourselves
This is where money should go, and where Denmark has the most room to act:
1. Shared security monitoring and funded support for small water and district-heating utilities without a security department.
2. Requirements for critical facilities to operate manually for several days, and for remote access to be closed or controlled.
3. AI on the defence side. Attackers already use it.
4. Shared requirements for critical infrastructure through NIS2 and EU cooperation to detect misuse of open-weight models.
5. Preparedness and credible crisis communication when rumours emerge.
Open question: Who pays for security at a small cooperative water utility? Nobody has answered that question properly yet.
6. Who should be able to say stop?
My proposal is independent safety testing before the most capable systems are rolled out, and a real possibility of stopping them. Denmark cannot do it alone. It must happen through the EU, where access to a market of around 450 million people provides leverage.
The USA has chosen another path. On September 29, the US government and leaders of the largest AI companies entered into a voluntary agreement on superintelligence safety. It is not legally binding and has no sanctions. Luiza Jarovsky points to unclear concepts and companies choosing their own auditors. In practice, this is self-regulation, which makes the EU's binding rules a real difference.
Testing before rollout is no guarantee that a system can be switched off after a loss of control. It is the best lever we have while it is still early.
7. Three things to take away
1. Measure impact and oversight. The government aims to free up at least 30,000 public-sector full-time equivalents through AI. Oversight, access to complaints and documented impact must accompany every krone in the acceleration fund.
2. Influence the EU now. The AI Act's high-risk rules have been postponed to December 2027 and August 2028. This is time to prepare oversight in Denmark and demand standards for how models are designed and tested.
3. Make transition a political priority. Skills development for small businesses and a career-transition fund that keeps pace.
These recommendations are my own, not the party's.
Status as of October 2026
Article 50 of the AI Act on labelling and transparency applies from August 2026. High-risk rules have been postponed to December 2027 and August 2028. The threat of destructive cyberattacks is high (raised September 24, 2026). This is political analysis, not legal advice: confirm dates and status against primary sources.
A question for you: Who in your organisation would notice if your systems were being mapped today?
About the podcast
The podcast version on Substack is AI-generated from my own presentation material. The voices are synthetic, and none of them is me. It opens with one of three fictional scenarios from the talk, without technical details. AI-generated audio can distort figures and wording. When the podcast and text disagree, the text and sources take precedence.
Sources
Status as of October 2026.
- International AI Safety Report 2026, February 2026, sections 2.1.2, 2.1.4, 2.2.2 and 3.4
- Geoffrey Hinton, Nobel banquet speech, December 10, 2024
- Morris et al., Levels of AGI, 2024, revised 2025, and Hendrycks et al., A Definition of AGI, 2025
- Pew Research Center, As Partisan Hostility Grows, August 9, 2022
- Regulation (EU) 2022/2065 (DSA), Articles 34–35
- Regulation (EU) 2024/1689 (AI Act), Articles 50 and 113
- SektorCERT, Angrebet mod dansk, kritisk infrastruktur, November 2023
- Danish Defence Intelligence Service, December 18, 2025
- Danish Agency for Societal Security, Truslen fra destruktive cyberangreb hæves til høj, September 24, 2026
- Anthropic, Disrupting the first reported AI-orchestrated cyber espionage campaign, November 2025
- Lambert and Brand, The ATOM Report, April 2026
- Directive (EU) 2022/2555 (NIS2)
- Det politiske grundlag for firkløverregeringen, June 2026, chapter 12
- Forbes (Sara Dorn) and Luiza Jarovsky, Luiza's Newsletter, both September 30, 2026
Originally published on Substack on October 5, 2026.
Disclaimer: This is political analysis and general guidance, not legal advice, and I am not a lawyer or compliance specialist. Do not use it as the sole basis for a decision, and confirm dates and status against primary sources (EUR-Lex, retsinformation.dk, the Danish Agency for Societal Security). These recommendations are my own, not the party's. Status as of October 2026.