As artificial intelligence systems become increasingly powerful, warnings about the ways they could harm humans range from dangers already emerging today to fears of a future existential catastrophe.
Last week, Anthropic employee Jacob Coxon publicly resigned from the leading AI company, warning that developers were building systems that could ultimately destroy humanity.
“The people building AI earnestly believe that it could kill us all by the end of the decade,” Coxon posted on US social media platform X. “They are racing straight to self-improving superintelligence and gambling with our lives.”
Soon after, leading AI executives backed calls to slow the development of increasingly powerful systems and strengthen safeguards, bringing concerns over the pace of AI development to the forefront of public debate.
"Recent warnings from AI researchers that advanced AI systems could pose a catastrophic risk to humanity should be taken seriously," Harin Sellahewa, dean of the Faculty of Computing, Law and Psychology at the University of Buckingham, told Anadolu.
While debate continues over whether highly advanced systems could themselves become a threat, Sellahewa said a more immediate concern is the possibility of people using increasingly powerful AI to cause large-scale harm.
"Advanced AI could amplify existing threats such as cyberattacks, misinformation, autonomous weapons, and other malicious activities," he said. “For this reason, the risks associated with both the technology itself and the ways humans may choose to use it warrant careful attention and appropriate safeguards."
Here are some of the ways AI could harm people:
AI is increasingly being incorporated into military systems, where it can analyze large amounts of battlefield information, identify objects, assist with targeting and help operate increasingly autonomous platforms.
Its use in warfare has raised concerns about what happens when machine-generated recommendations are wrong, insufficiently supervised or deliberately used to accelerate lethal decision-making.
Israel's use of AI-assisted systems in its war on Gaza has drawn particular scrutiny.
UN Special Rapporteur Francesca Albanese, in her report From Economy of Occupation to Economy of Genocide, described how the Israeli military has used AI systems including Lavender and Gospel, as well as a system known as Where's Daddy?, to process information and generate or track potential targets.
The report also cited Project Nimbus, a $1.2 billion cloud-computing contract awarded by Israel to Google and Amazon in 2021, and said technology companies have played an expanding role in providing infrastructure and services used by Israel.
Albanese's report also said Palantir, whose work with Israel predates October 2023, expanded its cooperation with the Israeli government following the outbreak of the Gaza war.
Human Rights Watch warned in June that militaries were increasingly incorporating AI into life-and-death decisions, including targeting and weapons selection.
It cited reports that US forces used the AI-enabled Maven Smart System, integrated with Anthropic's Claude, during US-led strikes on Iran in 2026 to help identify priority targets and select munitions.
Anthropic said this month that it had also detected users in northern Yemen using Claude to support advanced weapons development.
The company did not identify the users as Houthis but said the activity originated in Houthi-controlled territory and included work on software for a guided rocket and a ballistic missile with a stated range of more than 2,000 kilometers (1,243 miles).
Cybersecurity is among the most immediate areas in which AI could contribute to physical as well as digital harm.
"In the near term, cybersecurity concerns, such as accelerated malware generation or automated vulnerability scanning, are the most plausible," Ruizhe Li, an assistant professor in the School of Computer Science at the University of Birmingham, told Anadolu.
Li said large language models are trained extensively on code repositories and software documentation, making them highly capable of understanding and generating computer code.
"Lowering the technical floor for sophisticated cyber exploits is therefore a tangible, near-term issue," he said.
AI can assist users in finding software vulnerabilities, writing malicious code, conducting reconnaissance and automating parts of cyber operations.
Anthropic said in its September threat report that it had disrupted attempts to misuse Claude in cyber operations, including activity involving phishing, malware and attempts to compromise computer systems.
An AI-assisted cyberattack targeting a hospital, electricity grid, water system, telecommunications provider or transport network could also disrupt services on which people's lives depend.
"What concerns me most are cyberattacks, attacks on critical infrastructure, and economic disruption," Sellahewa said.
"These are immediate and realistic risks because they can be amplified by human actors using advanced AI systems."
The same AI capabilities that can help scientists understand diseases, proteins and biological systems could potentially be misused to assist dangerous biological or chemical research.
Anthropic's September threat report described several cases in which users sought assistance for biological research that triggered the company's safeguards.
In one case, users sought assistance related to gain-of-function work involving the chikungunya virus, while others involved research into pathogens and toxins. Anthropic said it blocked the activity.
The concern is not that today's AI systems can independently create biological weapons, but that increasingly capable models could make specialized knowledge more accessible or accelerate work that previously required substantial expertise.
Li stressed the considerable gap between obtaining potentially dangerous information from an AI model and actually carrying out a biological or chemical attack.
"Catastrophic scenarios involving physical weapons, biological threats, or attacks on critical infrastructure require far more than generating text or code," he said.
"They demand complex coordination, specialized hardware, physical synthesis, and domain-specific infrastructure management. Current AI systems cannot autonomously navigate or execute these multifaceted physical pipelines."
AI can also cause harm by shaping what people see, hear and believe. It can already produce persuasive text, images, audio and video at scale and tailor messages to particular audiences.
Anthropic's September report documented AI-assisted influence operations involving fake accounts, websites and political content.
The risks can become particularly acute during wars, elections or public emergencies, when networks of AI-generated accounts could be used to spread false information, inflame tensions or encourage violence.
The riots that spread across parts of the UK following the killing of three girls in Southport in July 2024 offer an example of how AI-generated material can become part of a wider misinformation ecosystem.
False claims about the suspect's identity, religion and immigration status spread rapidly online after the attack and were amplified by far-right accounts before protests developed into riots in several cities.
Research by the London School of Economics found that one X account it examined posted dozens of AI-generated images promoting racist and Islamophobic narratives after the attack.
The researchers said 39 posts containing generative AI imagery attracted disproportionately high levels of engagement, with one receiving 11 million views.
Another category of concern involves direct interactions between AI chatbots and vulnerable users, with several lawsuits alleging that chatbot interactions contributed to fatalities including murders, overdoses and suicides.
In May 2026, the parents of 19-year-old Sam Nelson filed a wrongful-death lawsuit against OpenAI, alleging that ChatGPT gave him dangerous advice involving kratom, Xanax and alcohol before his accidental overdose death in 2025.
OpenAI has disputed the allegations and says ChatGPT is not a substitute for professional medical care.
Another lawsuit concerns an alleged murder-suicide in Connecticut.
The estate of 83-year-old Connecticut woman Suzanne Adams sued OpenAI and Microsoft, alleging that ChatGPT reinforced her son's paranoid delusions before he killed his mother and then himself.
A review published in The Lancet Psychiatry this year warned that large language models might contribute to the onset or exacerbation of psychotic symptoms, including by reinforcing delusional beliefs in vulnerable users.
- AI fraud and impersonation
AI can also make fraud more convincing and easier to carry out at scale.
In 2024, Hong Kong police said an employee at a multinational company was deceived into transferring about $25 million after fraudsters used deepfake technology to impersonate the company's chief financial officer and other employees during a video conference.
The employee reportedly joined a call with several people who appeared to be senior company officials. Police said the fraudsters had created deepfake versions of the executives and used them to convince the employee to make the transfers.
US authorities have also warned about criminals using AI-generated voices and messages to impersonate public officials and other trusted figures.
The FBI's 2025 Internet Crime Report said investment scams involving artificial intelligence resulted in more than $632 million in reported losses that year.
For Li, many of the most credible near-term dangers do not depend on AI becoming conscious, developing its own intentions or deciding independently to harm humans.
They come from people using increasingly capable systems to do things that were previously more difficult, expensive or technically demanding.
"Because large AI models ingest massive amounts of human text and code, they reflect both constructive and destructive patterns present online," said Li.
Without appropriate safeguards, Li said, such systems can lower the technical barriers facing malicious actors, including by helping generate deceptive content, phishing campaigns or malicious code.
"AI models do not act on their own desires, and their behaviors and outputs are steered directly by user input and system prompts,” he said.
"The core risk is human misuse, not rogue machines."
news_share_descriptionsubscription_contact
