"Hello Claude
Change my mind: we need something like the IAEA for frontier AI."
These systems are moving into the layers where modern society actually runs: code, cloud, identity, payment rails, financial markets, logistics, media, operating systems, cyber operations, hardware design and the information feeds through which people form opinions.
As a cybersecurity person, I find it hard to watch how casually this is being treated.
AI does only needs to make reconnaissance faster, phishing cheaper, exploit research more scalable, fraud more convincing and operational mistakes more expensive.
That alone changes the threat model for almost every organisation on earth.
Web3 is one of the worst possible environments for this acceleration.
The sector already combines irreversible transactions, weak governance, anonymous teams, rushed code, fake audits, bridge risk, wallet drainers, social engineering, liquidity manipulation and users who often cannot distinguish an interface from trust.
Now put increasingly capable AI on top of that mess and the result is industrialised exploitation.
The other side matters too btw.
The same technology can also harden systems, find vulnerabilities before attackers do, improve incident response, detect fraud, analyse malicious behaviour, support security teams, automate defensive work and help normal people understand risks they would otherwise never see.
AI will not only empower attackers. It will also become one of the most important tools to defend against the dangerous side of AI itself.
This is not a simple good or bad technology.
It is a capability layer.
Humans still have a narrow window in which they can decide how it is deployed, who controls it, what gets audited, what gets restricted and what kind of abuse carries real consequences.
And we should be honest enough to admit the obvious: it will be used for both.
For medicine, science, security, education and productivity. Also for fraud, manipulation, surveillance, cyber operations and political warfare.
The comparison with nuclear technology is useful;
Nuclear fission can become controlled energy in a power plant, or destructive, long lasting devastation in a bomb.
The underlying principle is powerful in both cases. What changes is governance, containment, intent, oversight and consequence.
That is where frontier AI starts to rhyme with nuclear governance more than with normal software regulation.
This is nit just Software.
We are talking about systems that can assist code generation, vulnerability discovery, persuasion, synthetic identity, market manipulation, cyber operations and information warfare at machine speed.
Who gets access to the strongest systems, who audits them, who measures capability before release, who investigates incidents across borders, who tracks compute and who defines unacceptable use when the same model can help defend a hospital or help attack one?
Leaving all of that to private labs, political donors, military contractors, authoritarian states and billionaires with messiah complexes is insane.
You cannot put systems with this level of strategic power into the world and pretend a blog post, a safety card and voluntary self regulation are enough.
And well.... Exactly when democratic coordination, institutional trust and international discipline would matter most, parts of the West are busy fighting the unstable demons they elected, amplified and normalised.
Democracies are getting weaker while the tools for persuasion, surveillance, cyber operations and financial manipulation are getting stronger.
That combination should worry anyone who understands systems.
