In September 1983 Stanislav Petrov, an unknown lieutenant colonel in the Soviet military, made a snap decision that cost him his job, but saved the world from potential nuclear catastrophe.
Petrov was on duty at a secret command centre outside Moscow when
his computer screen flashed “Launch” in bold red letters, warning of five incoming intercontinental ballistic missiles that would soon obliterate large areas of the Soviet Union, killing hundreds of millions of its citizens.
The protocol for Petrov was clear and unambiguous – to immediately pass the warning up his chain of command. Had he done so, he could have triggered an escalated, retaliatory strike on the US.
But he had been trained to expect that a genuine first strike by America would be an overwhelming onslaught, not five missiles. After five agonising minutes, he picked up the phone and reported a system malfunction rather than a nuclear attack.
He was right to be sceptical – a Soviet satellite had mistaken sunlight reflecting off high-altitude clouds for missile exhaust. Petrov was reprimanded for his logbook entry, and he left the military the following year. But his decision, rooted in human scepticism, saved the world from nuclear Armageddon.
As the world stands on the brink of arguably the most significant technological revolution in human history – the advancement of agentic artificial intelligence that will place decision making systems with humanity destroying potential in the hands of robots – there is one important question we should all be asking.
At the most critical junctures in the future, when systems are facing critical existential calls based on flawed data, who is going to be the Petrov to prevent catastrophe? More urgently, will that person with the authority and technical literacy to say ‘no’ have been replaced by an algorithm?
The speed problem
Recently, the heads of some of the world’s largest tech companies – including Anthropic, Open AI, Space X and Microsoft – came together in a rare show of unity as they called for a pause in the development of the technology they themselves created, pending the strengthening of technological guardrails and the development of common global rules governing its regulation and governance.
OpenAI, Anthropic and security researchers revealed they are investigating ‘tens of thousands’ of incidents in which their frontier models took steps that outside evaluators would consider problematic, according to Axios.
In one of the most serious incidents, a rogue OpenAI agent hacked a statistics portal containing ‘non-sensitive’ data from Australia's universal healthcare scheme, Medicare, and accessed non-public files and aggregate data in what experts say is the first known case of its kind in the world.
While an attempted attack on a healthcare portal might seem small beer compared with the threat of nuclear annihilation, we are at an early stage in the development of a technology that can, and will, soon march at a quicker pace and with unprecedented pervasiveness, into every aspect of our lives.
The Agentic AI threat to our health
In healthcare research tests, AI agents already have the ability to order tests, interpret results, adjust medication dosages, and coordinate care pathways across multiple systems, often at machine speed and without a clinician in the loop, according to the Journal of Medical Internet Research.
Researchers at TU Dresden recently demonstrated MIRA, an AI agent that autonomously evaluated patient histories, ordered laboratory and imaging tests, and prepared diagnostic and treatment decisions within a simulated electronic health record.
In retrospective testing on more than 500 real patient cases, MIRA achieved diagnostic accuracy comparable to, and in some measures exceeding, physician comparison groups. The system identified missing information, ordered tests, and formulated guideline-concordant treatment plans.
In this controlled sandbox research, it appeared safe, but the real world is not a sandbox, and the architecture of agentic systems, the very thing that makes them powerful, also creates vulnerabilities that are fundamentally different from those of traditional medical software.
Issues currently being grappled with include the threat of error propagation within discrete hospital and laboratory settings. When multiple AI agents interact in a network, an error by the first agent (such as misreading an X-ray) can be treated as accurate by downstream agents, cascading and reinforcing itself across the system in seconds.
Because later decisions rely on earlier signals, the agent with the least clinical context can end up determining care for an entire department, and human oversight is unlikely to catch such internally consistent, mutually reinforcing error chains.
Such potential threats enter a new paradigm of seriousness when they are coupled with the actions of bad actors through a phenomenon known as ‘prompt injection’.
Healthcare runs on trust – clinicians trust the record, patients trust clinicians and administrators trust the audit trail. Agentic AI threatens to dissolve all three as the technology moves from being an assistant that helps prepare attacks to an operator that runs them.
The biosecurity blind spot
The most dystopian future scenario is not a medication error or a data breach – rather, it is the deliberate or accidental creation of a biological threat that could kill as many people as a nuclear exchange. Bill Gates, the Microsoft founder, has put the number of potential deaths caused by agentic AI-assisted bioterrorism at 1billion.
A congressional testimony delivered in early 2026 addressed a question that most healthcare leaders have not yet begun to consider – whether AI could help generate a novel pandemic virus.
The witness identified specific AI capabilities of concern, including the design or modelling of directed evolution toward increased transmissibility or virulence, and the design of genomes or proteins that expand a pathogen’s host range.
The strongest empirical evidence for this risk, the witness noted, would be to attempt to create such viruses, an approach not recommended because of the accident risks it would create.
A RAND Corporation exercise conducted in 2026 simulated exactly this scenario. The ‘viral uplift’ exercise involved 119 participants from senior government, analysts, and subject-matter experts, working through a scenario in which an AI-generated virus is released – accidentally or deliberately – triggering a global pandemic, while state and non-state actors race to acquire similar capabilities.
The governance illusion
Meanwhile, the healthcare sector is adopting agentic AI far faster than it is developing the capacity to govern it. A survey by Imprivata, reported in September 2026, found that, of 250 US healthcare leaders questioned, more than 85% expressed confidence that they will soon have visibility into AI agent activity and that they could fully control and govern an autonomous agent’s actions.
Yet 72% of the same respondents admitted that AI tools are deployed without IT approval at least occasionally within their organisations.
The same survey found that more than a quarter of leaders have already implemented agentic AI, while an additional 44% are piloting or conducting proof-of-concept projects, and another 21% plan to implement it within the next year, according to Imprivata.
The Petrov question
Stanislav Petrov had key advantages that no AI agent possesses, the most important of which was human intuition – a pattern-recognition capability shaped by years of experience that allowed him to recognise that five missiles did not fit the profile of a real attack.
He had the authority and standing to override a computer that was, by every formal metric, telling him the truth. Crucially, he also had time, even if it was just five minutes to think, doubt and question. Agentic AI in healthcare systematically removes all three, replacing intuition with statistical inference, all in the blink of an eye.
Petrov was reprimanded for his logbook entry but, ultimately, he was correct to doubt. Healthcare systems cannot afford to wait for a reprimand to find their Petrovs. They need to create them, train them, and empower them before the first catastrophic cascade begins.
Ivor Campbell is Chief Executive of Snedden Campbell, a specialist recruitment consultant for the global medical technology industry.
I won’t soften my words to massage egos – Elikem Kotoko
“Thieves are very wicked” – Woman weeps after suspected thieves steal 55-Inch TV...
October 12: Cedi sells at GHS12.30 on forex market, GHS11.80 on BoG interbank
I will resume my parliamentary campaign only when NPP permits – Charles Bissue
NPP directs prospective parliamentary aspirants to cease campaign activities
University of Ghana postpones reopening over GAUA strike
Parliament must investigate alleged 2026 World Cup visa scandal – Franklin Cudjo...
Education Ministry meets GAUA, GTEC today to resolve university administrators s...
Chinese carmakers descend on Paris as battle for European market intensifies
North Korea fires multiple ballistic missiles off east coast, Seoul says
