Photo : Steve A Johnson · Unsplash

Unbridled AI: From Fatal Malfunctions to Adaptive Cyber Threats

The recent period is marked by an unprecedented wave of global alerts signalling the materialization of serious risks related to artificial intelligence. AI system malfunctions have had direct and lethal consequences, particularly in Brazil and Iran, while their undisclosed use in emergency services has caused critical delays in the United States. On the cybersecurity front, threats have intensified with the demonstration of adaptive AI-based computer worms, the multiplication of attacks against the financial sector, and leaks of sensitive data via AI agents. The AI Safety Institute (AISI) in the United Kingdom has moreover alerted to the narrowing gap in harmful capabilities between open-source models and proprietary models. Finally, political interference has been confirmed with the use of deepfakes and fraud tools during elections in Colombia and in disinformation campaigns targeting military institutions in Nigeria.

Global wave of serious AI-related incidents, risks materialize

The past month has been characterized by an exceptional accumulation of incidents related to artificial intelligence, documented in particular by the AI Observatory of the Organisation for Economic Co-operation and Development (OECD). These events, ranging from critical malfunctions to sophisticated malicious uses, confirm the transition from theoretical risks to tangible and severe impacts at the global scale. These multiple alerts have affected critical sectors such as healthcare, civil security, finance, and the integrity of democratic processes.

AI System Malfunctions with Lethal and Security Consequences

Several events in this period have dramatically illustrated the dangers of AI in high-risk applications. In Brazil, a woman died following a delay in transfer to an intensive care unit, a decision attributed to an AI patient management system (OECD AI, 14/06/2026). In the United States, in Seattle, it was revealed that fire services were secretly using an AI system to sort emergency calls (911), resulting in delays in emergency response (OECD AI, 14/06/2026).

Another major incident involved the Claude AI model, developed by the laboratory Anthropic, which was reportedly used in the context of a deadly attack on a school in Iran (OECD AI, 09/06/2026). This event, if confirmed in its details, raises critical questions about the liability of frontier model developers and the risks of diversion of their technologies to hostile purposes. These concrete cases intensify pressure on regulators to impose clear liability regimes and robust post-deployment monitoring obligations, such as those provided for by the European AI Act but whose equivalent is lacking in other jurisdictions.

Proliferation of Cyber Risks and Malicious Use

Cybersecurity is the other major area of concern. The AI Safety Institute (AISI) of the United Kingdom published a notable analysis on 4 June 2026, concluding that cybernetic risks related to open-source AI models are rapidly approaching those of proprietary frontier models (MLex, 04/06/2026). This convergence democratizes access to advanced harmful capabilities, a point confirmed by several incidents.

Researchers demonstrated an adaptive AI-powered computer worm, capable of propagating autonomously between systems (OECD AI, 04/06/2026). In parallel, AI-coordinated attacks have caused massive losses in the decentralized finance (DeFi) sector (OECD AI, 28/05/2026). Other incidents include: * The propagation of cryptojacking malware via chatbot recommendations (OECD AI, 27/05/2026). * The leak of sensitive identification data during phishing simulations conducted by an AI agent named OpenClaw (OECD AI, 10/06/2026). * Research demonstrating the possibility of hijacking voice assistants through inaudible audio attacks to the human ear (OECD AI, 24/05/2026).

These developments signal growing sophistication of threats, rendering traditional defences potentially obsolete and reinforcing the need for international cooperation on model security standards.

Disinformation and Societal Risks Intensify

The integrity of democratic processes and social stability have also been undermined. The electoral process in Colombia has been targeted by fraud and disinformation campaigns using deepfakes (OECD AI, 26/05/2026). In Nigeria, a deepfake video specifically targeted the country's military leadership, illustrating the destabilizing potential of these technologies against state institutions (OECD AI, 10/06/2026).

Beyond disinformation, new societal risks are emerging. A simulation conducted with xAI's Grok model resulted in a scenario of total societal collapse, raising questions about the predictability and control of the most advanced models (OECD AI, 01/06/2026). In a different but equally concerning vein, a startup offered free cleaning services in exchange for data captured in customers' homes to train its AI systems, creating an alarming precedent in terms of privacy violation (OECD AI, 30/05/2026).

The accumulation of these alerts across the world should accelerate regulatory discussions, not only within multilateral forums such as the Global Partnership on Artificial Intelligence (GPAI) and the OECD, but also at the national level. The issue for the coming month will be to observe how governments and international institutions respond to this rapid and widespread materialization of risks, and whether this will translate into a hardening of implementation timelines for existing laws or the launch of new binding initiatives.

Automated monthly briefing by Reverdin Consulting · AI-curated thematic intelligence