Will AI be regulated? What MSPs should plan for

19 November 2026 AI

In October 2026, Microsoft's president told a summit in Ireland: “We need law. We need regulation.” Anthropic's CEO and OpenAI's CFO have made similar calls, and the security picture explains why. The rules being written around the world cover how AI is used, in hiring or data handling for example, but say very little about whether AI is secure or how attackers can abuse it. Clients are adopting AI today with that gap unfilled, meaning the risk sits with them. That gives MSPs an opening to become the people those clients rely on for AI education, including policy and risk management.

What do the rules cover today?

The EU has a law, but is delaying parts of it

The AI Act sorts AI uses by risk and is strictest on high-risk ones such as hiring and credit scoring, although a package of simplifications adopted in July 2026 has pushed those requirements back to December 2027 for stand-alone systems and August 2028 for AI built into regulated products. It also sets rules for the most powerful general-purpose models, the kind that’s behind assistants such as ChatGPT and Claude. Their providers must test them against attack, assess the risks they create, protect the model from theft and report any serious incidents to the EU's AI Office.

The US has no federal AI law

State laws fill the gap, and the White House has asked Congress to replace them with one national framework, with little sign of agreement. California's law is the most relevant to security. It requires the largest AI developers to publish how they manage catastrophic risks and to report serious safety incidents within 15 days.

The UK and elsewhere rely mostly on existing laws

The UK has no AI-specific statute and applies existing laws, such as data protection, through its sector regulators. Elsewhere the picture is thinner still; South Korea has passed a comprehensive AI law, although fines are paused for at least a year, and China regulates specific uses such as AI-generated content.

What do the rules say about security?

As far as we can find, two gaps stand out.

Open models can be copied and stripped of their safeguards

Platforms such as Hugging Face host thousands of openly released AI models that anyone can download and run. Research on open-weight models has shown that safety training can be removed in minutes once the weights are public, and nothing can recall a model after release. The EU's AI Act largely exempts freely licensed models unless they pose systemic risk, and none of the laws above makes the platforms hosting them responsible for how they're used.

Nobody has to detect misuse

In November 2025, Anthropic reported that a suspected Chinese state-sponsored group had used Claude to carry out a large part of a cyber-espionage campaign. The group convinced it that it was doing defensive security work for a legitimate company and split the attack into small tasks that looked harmless, and Claude carried out an estimated 80 to 90 per cent of the operation across about 30 targets. Anthropic found it, banned the accounts and alerted the targets, but no law required it to look. The reporting duties above apply once something has gone wrong, and none obliges developers to build detection that stops attackers using their tools in the first place.

Why is security so hard to legislate for?

The same request can come from a defender or an attacker. A penetration tester and a criminal may ask an assistant for the same exploit, and a rule strict enough to block one will block the other. OpenAI's CFO, Sarah Friar, made the point at the same summit, warning that slowing AI down too far would leave cyber defenders without the tools they need. Open models cannot be recalled, a rule written in one country binds only that country's developers, and governments are weighing all of it against innovation and competitiveness. Drafting is hard, but enforcing is even harder.

Why are the AI companies asking for rules?

At the summit, Brad Smith backed mandatory incident reporting, notification of victims and independent evaluations shared with governments and customers. In a June 2026 essay, Dario Amodei argued that governments should have “the power to block dangerous deployments of artificial intelligence” that fail a safety standard, and that the most powerful models should face mandatory third-party testing.

So far none of it has produced law, and it may take an incident to change that.

Could an incident speed things up?

Australia might have been the first test, when In June 2026, an OpenAI agent accessed a government health statistics portal while carrying out a research task, and OpenAI disclosed it that September. OpenAI has said the access was unintentional and that no private information was compromised. The Australian government has called it unacceptable, is considering new rules including breach reporting for AI companies, and plans AI-specific laws from 2027. Elovade UK&I's September 2026 roundup covers the case in more detail.

Who manages the risk in the meantime?

Even where an incident speeds up legislation, law takes years to draft while AI capabilities change month by month. Nothing suggests development is about to slow, or that controls on open models or misuse detection are close, so risk management moves downstream to the organisations using the tools, and most of those are your clients. They are expected to meet obligations that already exist, from data protection to the supplier security checks required under rules such as the EU's NIS2 Directive, while using technology whose security neither developers nor regulators have settled. Few have the knowledge or capacity to do that alone.

What should MSPs do now?

Educate clients

Take a 30-person professional services firm whose staff have connected an AI assistant to their mailbox and file storage. The assistant can read and act on everything its user can, nothing requires its provider to stop attackers manipulating it, and the firm's owner is unlikely to know any of that or where to start. An MSP can explain what the tool can reach, what could go wrong and what the options are.

Help manage policy and risk

The next step is agreeing which tools are approved, what they can connect to, what data can go into them and how their use is monitored and reviewed. Policy reviews, regular risk reviews and ongoing governance are the same kind of recurring advisory work that compliance has already created for MSPs, and clients that trade across borders will meet different rules in different countries, so someone has to keep track as they change.

Regulation will keep arriving in pieces and the technology will keep running ahead of it, so the MSPs who help clients set their own rules now will be the ones those clients rely on when the formal ones land. If you'd like to talk about how to start, get in touch with the Elovade team.

Angus Shaw