The security of cutting-edge AI systems is now at the forefront of regulatory concerns, marked by a major alert from the UK's AI Safety Institute (AISI) on rising cyber risks (AI Safety Institute — News (via Google News)). This analysis is corroborated by the security incident of June 5, 2026, when malicious actors exploited Meta's AI-based customer support agent to steal Instagram accounts (MIT Tech Review). In parallel, the Anthropic laboratory announced that its Mythos model was too proficient in hacking techniques to be made public (MIT Tech Review) and called for a global slowdown in AI development (MIT Tech Review). On the multilateral front, the Organisation for Economic Co-operation and Development (OECD) launched on June 5, 2026 a new AI policy toolkit for governments (OECD AI / GPAI — News (via Google News)), while also acknowledging that for 40% of employers, the lack of skills is the main obstacle to AI adoption (OECD AI / GPAI — News (via Google News)). Finally, a weak signal with high societal impact is emerging from the research of Gloria Mark, who fears that digital technologies, and now AI tools like ChatGPT and Claude, are weakening our cognitive abilities (MIT Tech Review), drawing on data showing a drastic decline in average attention span (MIT Tech Review).
Cutting-edge Models: The Materialization of Security Risks
The monitoring period was marked by the realization of security risks related to AI, moving from the theoretical field to proven incidents.
The hijacking of Meta's AI agent constitutes a turning point. On June 5, 2026, it was reported that attackers used Meta's conversational agent dedicated to customer support to take control of Instagram accounts (MIT Tech Review). This incident illustrates the vulnerability of AI systems in production, even those that are not foundation models, and opens a new attack surface for companies deploying autonomous agents in interaction with the public.
This practical alert directly echoes the strategic warnings issued by regulators and developers. The UK's AI Safety Institute (AISI) published an analysis on June 4, 2026, concluding that cybernetic risks are intensifying as open-source AI closes its gap with proprietary foundation models (AI Safety Institute — News (via Google News)). This convergence of capabilities makes cyber-offensive tools potentially more accessible, a risk no longer confined to state actors or cutting-edge laboratories.
The Anthropic laboratory, one of the leading developers of foundation models, has reinforced this perception of risk. On one hand, it announced that its Mythos model proved to be too proficient in hacking techniques to allow widespread distribution (MIT Tech Review). On the other hand, the company called for a general slowdown in the race for AI development on a global scale (MIT Tech Review), and explicitly flagged the risk that models could become capable of "self-improving" autonomously (MIT Tech Review). This position, coming from a leading commercial player, is a strong political signal addressed to decision-makers on the urgency of robust oversight before model capabilities exceed control mechanisms.
The OECD Strengthens Its Tools for AI Governance
Faced with technological acceleration and associated risks, multilateral organizations continue their efforts to structure governance. The Organisation for Economic Co-operation and Development (OECD) launched on June 5, 2026 its AI Policy Toolkit, a set of tools designed to help governments implement public policies for AI (OECD AI / GPAI — News (via Google News)). This initiative aims to translate the OECD's high-level principles into practical and actionable instruments for member states, covering areas such as regulation, public investment, and skills development.
The emphasis on skills is not insignificant. A recent OECD study highlights that for 40% of employers, the main barrier to AI adoption is neither cost nor technology, but a skills deficit within the workforce (OECD AI / GPAI — News (via Google News)). This figure sheds light on a major strategic challenge for public policies: successful AI adoption will depend as much on training and adaptation of human capital as on regulatory frameworks.
The Deeper Impact on the Digital and Cognitive Ecosystem
Beyond security and governance issues, underlying trends raise questions about the impact of AI on society and individuals.
The work of Gloria Mark, a psychologist at the University of California, Irvine, warns of the weakening of our cognitive abilities by digital technologies (MIT Tech Review). Her research indicates that the average attention span has declined dramatically over the years (MIT Tech Review). It dropped from approximately two and a half minutes in 2003 (MIT Tech Review), to 75 seconds in 2012 (MIT Tech Review), to only 47 seconds on average between 2014 and 2020 (MIT Tech Review). Gloria Mark now estimates that AI tools like ChatGPT and Claude could accelerate this phenomenon (MIT Tech Review). According to her, "You delegate your cognitive work to AI. And that's not good for us" (MIT Tech Review). This perspective questions the very design of generative AI applications and their long-term impact on attention, memory, and critical thinking capacity.
In parallel, the composition of internet traffic itself has crossed a symbolic threshold. Cybersecurity company Cloudflare indicated that 57.4% of global web traffic now comes from bots, rather than humans (MIT Tech Review). This shift surprised even Cloudflare co-founder and CEO Matthew Prince, who commented: "Well, that happened faster than I predicted" (MIT Tech Review). This "botification" of the web, accelerated by AI agents, has profound implications for cybersecurity, the information economy, audience measurement, and the very nature of digital public space.
Photo: Igor Omilaev / Unsplash