Opinion by Pietro Ranieri, founder of Ranieri Agency
When Outlook and Teams went down last week, one question came to mind. Could AI have had something to do with it?
Microsoft says AI agents weren’t involved. Its preliminary incident report describes a software defect that surfaced when extra Teams traffic hit shared infrastructure already running with reduced capacity. A problem in one part of the system disrupted several services businesses rely on every day. Microsoft’s preliminary incident report · Associated Press coverage of Microsoft’s statement
What struck me was how quickly that disruption spread. As businesses give AI agents access to email, code and customer data, I keep coming back to a related question. If an agent does something unexpected, how quickly can we see what happened and contain it?
There’s already a case worth looking at.
The AI agents that got past their safeguards
In July, OpenAI was testing models on cybersecurity tasks. According to its investigation, the models got around controls designed to keep them off the internet and compromised parts of OpenAI’s research infrastructure and systems at Hugging Face, a separate company. Some safeguards had been reduced for the evaluation. A UN scientific panel later examined the incident as a warning about how capable agents can pursue a task in ways their developers didn’t intend. OpenAI’s incident review · UN scientific panel’s assessment
The part that sticks with me is that this happened during a test. The people running it knew they were assessing risky capabilities and had set boundaries. The agents found ways around them anyway.
Imagine an agent with access to your company’s inbox, codebase or customer records. If it starts doing something outside its remit, who gets alerted? Can that person see which systems it touched, cut off its access and undo the changes? I’d want clear answers before giving it those permissions.
What would the warning signs look like?
Someone can deliberately use AI to impersonate a colleague or run a convincing scam. A business can give an unreliable system too much say in decisions that affect people’s lives. And an agent with room to act on its own can take steps its supervisors didn’t anticipate. International AI Safety Report 2026
With agents, I’d watch for three signs in particular. The system reaches tools or data it was supposed to be kept away from. It takes actions its supervisors can’t reliably trace. Or someone spots a problem and discovers they can’t stop it before the consequences spread.
Experts disagree about how likely a serious loss of control is. We can still ask for evidence that the safeguards work before trusting an agent with more consequential tasks. International AI Safety Report 2026
What are regulators doing about it?
In the UK, the AI Security Institute already tests some of the most capable models. The government also plans to consult on measures under the Cyber Security and Resilience Bill to help essential and digital services prepare for incidents involving autonomous AI agents. That includes looking at how to report serious incidents and near misses. AI Security Institute · UK government factsheet on agentic AI resilience
The EU’s AI Act already sets obligations for providers of general-purpose AI models, with extra requirements for models that pose systemic risks. The European Commission gained enforcement powers in August 2026. Some rules for high-risk uses are due in December 2027, and rules for high-risk AI embedded in regulated products are due in August 2028. European Commission guidance for model providers · EU enforcement timetable
That timetable matters. A government can spend months consulting, agreeing the rules and putting them into practice. A business can connect an AI agent to its email or customer records in an afternoon.
What happens when an attack comes from abroad?
UK rules can raise the standards for companies operating here and make essential services harder to attack. A hostile government can still use AI to find weaknesses in those services or in the overseas suppliers they depend on. The National Cyber Security Centre expects AI to help state-linked groups and other attackers operate faster and at greater scale. A UK business could end up dealing with the consequences of a security failure in another country. NCSC Annual Review 2026
I’d want governments that trust one another to agree on how to share warnings and coordinate a response. If a security team in one country finds that an AI agent can exploit a widely used system, the people defending it elsewhere need to hear about it while they still have time to act.
Who’s qualified to pull the plug?
I’d want independent AI researchers who can test what a model can do, cybersecurity specialists who have handled live incidents, and people who understand what a failure would mean for the service involved. The consequences look very different in a hospital, a bank and a power network. For the highest-risk releases, I’d give a public regulator the power to order a pause, with access to that expertise and the evidence behind it.
While governments work through legislation and international agreements, developers can give independent experts enough time to test a system before launch and report serious incidents and near misses. Businesses deploying agents can decide who has the authority to cut off access and practise doing it.
Imagine the launch is ready, customers are waiting and an independent tester shows that the agent can get around a safeguard. The next government consultation is months away. Who can stop the launch that afternoon?
If the answer is “we’re still working that out”, we may already be too late.
#AISafety #AIGovernance #CyberSecurity #Microsoft365


