A sharp sell-off in artificial intelligence-linked stocks on Monday was not a sign that the "AI bubble" had burst, nor a sudden realisation that artificial intelligence is dangerous, according to Capital.gr. Instead, investors reacted to public agreement among Dario Amodei, Sam Altman and Elon Musk, three of the most influential figures in AI, that development of the most powerful models needs to proceed at a pace that allows safety mechanisms to catch up.
Investors did not wait for detailed explanations. They sold first and read the fine print afterwards, according to the report.
The nervousness had already been fuelled by oil prices that climbed above $108 a barrel during the session and by the yield on the US 10-year Treasury note, which temporarily crossed 5%. The comments on AI acted as a trigger rather than the sole cause of the sell-off.
The index tracking US semiconductor stocks lost almost 6%, while the Nasdaq fell by around half a percentage point. Microsoft, Alphabet and Meta all closed higher, showing there was no broad exit from AI-related shares. The market instead punished companies most reliant on continued spending on chips, data centres and electricity. A slowdown in that race would hit infrastructure suppliers first, though demand for computing power itself is unlikely to disappear; it may simply shift from training ever-larger models to running the ones that already exist.
What Amodei, Altman and Musk said
What stood out was that three executives who compete fiercely, and in the case of Musk and Altman by almost every means available, found common ground on something as fundamental as the pace of development.
Amodei, chief executive of Anthropic, gave the most comprehensive intervention, addressing cyberattacks, biological risks, economic disruption, loss of control and what he called recursive self-improvement, the growing ability of today's models to help build tomorrow's and so speed up the pace of development itself.
Altman, chief executive of OpenAI, publicly agreed that the pace of development of the most powerful models needs to be controlled and committed to bringing in independent evaluators. Musk limited his comment to saying that Amodei was right, a remark carrying particular weight given his years of warning about the dangers of AI.
Cybersecurity risk is already here
Not all the risks are the same, the piece argues. Cybersecurity risk is already present: systems have been recorded carrying out target reconnaissance, searching for vulnerabilities, helping steal passwords and organising attacks with far less human involvement than a few years ago. AI did not invent hackers, but it has given them an army of cheap, tireless and fast assistants. The same applies to scams, disinformation and mass surveillance. There is also a military dimension, as autonomous target selection, intelligence analysis, drones and faster decision-making can dramatically shrink the time a person has to decide on an action with irreversible consequences.
Biological risk, real but overstated
Biological and biochemical risk is serious but does not have the scale often attributed to it. Models can now meaningfully help draft laboratory procedures and solve difficult scientific problems, but the gap between that and any individual building a pandemic in a garage still requires laboratories, materials, equipment and genuine expertise. The issue needs careful handling, not hysteria.
Autonomy: the harder problem
More complex is the question of autonomy. The serious danger is not an AI that "wakes up," develops a will and decides it dislikes humans. It is a system with significant autonomy, access to tools and a poorly defined goal that keeps carrying out instructions in ways nobody predicted. Causing damage does not require consciousness, only capability, access and inadequate safeguards. By contrast, scenarios involving full autonomous self-replication or an AI "taking over" the internet remain purely theoretical.
Power concentration, a quieter risk
A less dramatic but probably more significant risk is the concentration of power. If a handful of labs control the models used by governments, businesses, banks, media organisations and military bodies, the issue stops being purely technological. It becomes a new form of economic and political infrastructure, comparable in importance to telecommunications or the financial system but concentrated in very few hands. The answer is not to stop large companies from growing but to keep the market genuinely open to competition, through common technical standards, the ability to move data and workloads between providers, and limits on practices that lock customers into a single platform. Competition itself functions as a safeguard.
Wages and who captures the gains
The economic risk is just as real. AI raises productivity and reduces the need for some jobs, particularly entry-level and repetitive office work. It has not caused mass unemployment. But it can compress wages, limit opportunities for young workers, weaken mid-level specialisation and shift an even larger share of wealth to a small number of companies. The social threat, the piece argues, is not only how many jobs disappear but who captures the productivity gains.
What could be done
Halting AI development altogether is not a realistic answer, being both impractical and geopolitically risky if rivals continue regardless. In the short term, the piece calls for independent evaluators, mandatory reporting of serious incidents and strict testing environments, along with genuine multi-layered authorisation involving a human, independent organisational oversight and technical limits separate from the model itself, rather than simply a second AI system approving the first.
In the medium term, it proposes automatic safeguards similar to stock market circuit breakers: when a new model exceeds set thresholds in cybersecurity capability, biological knowledge or autonomy, research would not stop, but external access would be temporarily restricted until further checks are completed. This should be paired with mandatory insurance, capital reserves against risk, and a shared fund for catastrophic losses financed entirely by the companies themselves, scaled to the level of risk their systems create, since extreme AI risks are exactly the kind insurers struggle most to price: rare, enormous and potentially simultaneous across many firms.
AI could also be used defensively, with specialised systems scanning networks, identifying vulnerabilities and suggesting fixes before attackers find them, though any active intervention in real systems should remain tightly limited and require human approval to avoid triggering a machine-versus-machine arms race.
In the long run, the most serious solution is verifiable international transparency rather than naive international agreement. Major powers could agree on common tests for cybersecurity, biological and military risks, and on technical mechanisms that prove training of a new powerful model stays within agreed limits without revealing code or trade secrets, akin to a "digital arms inspection" relying on cryptographic proof rather than an inspector with a file.
A convenient kind of caution
There is also another motive at play, the piece notes. Large players benefit from rules so precise and complex that smaller rivals cannot meet them, raising the barrier for the next competitor and reducing the enormous cost of the next generation of models. That does not mean Amodei, Altman and Musk are pretending to be afraid. It means something more interesting: safety and corporate interest can coincide, and when the wolves ask for a better fence, it is worth checking exactly where they want it built.
The market probably overreacted on Monday, the piece concludes, but not because there is no problem. It overreacted because it confused the need for safer development with the end of development altogether. What is genuinely new is not that the people building the most powerful models suddenly discovered risk. It is that they have begun publicly calling for containment mechanisms while taking part in the most expensive and aggressive technology race in history. If that shift moves from words to rules, it could change more than the valuation of a few stocks. It could change the economic model the entire AI boom has been built on.
