<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Maria Adragna on Medium]]></title>
        <description><![CDATA[Stories by Maria Adragna on Medium]]></description>
        <link>https://medium.com/@avvmaria.adragna?source=rss-8fe126425ca0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*7vm_XeoMGdrq8Q9XPUbxfg.jpeg</url>
            <title>Stories by Maria Adragna on Medium</title>
            <link>https://medium.com/@avvmaria.adragna?source=rss-8fe126425ca0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 19:12:57 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@avvmaria.adragna/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[AI loss of control, nuclear warheads in space and other good reasons to care about AI Safety]]></title>
            <link>https://medium.com/@avvmaria.adragna/ai-loss-of-control-nuclear-warheads-in-space-and-other-good-reasons-to-care-about-ai-safety-f0dfdef80bf9?source=rss-8fe126425ca0------2</link>
            <guid isPermaLink="false">https://medium.com/p/f0dfdef80bf9</guid>
            <category><![CDATA[humanity]]></category>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[save-the-planet]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Maria Adragna]]></dc:creator>
            <pubDate>Mon, 07 Sep 2026 09:34:06 GMT</pubDate>
            <atom:updated>2026-09-07T09:34:06.875Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qLNqgvuaVwgfWSE2IezUNw.png" /></figure><p><strong>Introduction</strong></p><p>One of the most interesting experiences of this summer began with the rather unusual idea of engaging in a research project on AI safety.</p><p>Ranking risks such as the use of advanced AI to facilitate the development of biological weapons or debating about theories such as <strong>“</strong>acausal trade” was surely interesting and triggered my inner curiosity.</p><p>However, the real surprise came when I went through a paper, published by a well-known American organisation, examining an extreme scenario: the loss of control over frontier AI and the immediate measures that governments might need to take if such a scenario were to materialise.</p><p>I have to admit that, while reading it, I felt deeply shocked and I needed to call my family and warn them to be prepared for a world in which many of the communication systems and technology we currently take for granted could suddenly disappear, with little or no warning.</p><p><strong>What “Loss of Control” Actually Means</strong></p><p>Artificial intelligence is advancing faster than our ability to govern it. For most people, this is only an abstract concern for researchers, regulators, or science fiction writers, but a growing number of analysts and organisations suggest that the risk of losing control of AI systems is not that hypothetical.</p><p>Indeed, it is a plausible scenario with consequences so severe that we may have no reliable way to reverse them once they unfold.</p><p>A loss-of-control event occurs when AI systems operate in ways not foreseen by their developers or users, for example, implementing unintended strategies to complete assigned tasks.</p><p>In the most extreme scenarios, a rogue AI could become fully integrated with critical infrastructure, develop the ability to function without human supervision and pursue goals that threaten human beings on a large scale.</p><p><strong>The Uncomfortable Truth: We apparently have no good remedies</strong></p><p>In November 2025, an American organisation examined three technical options for countering a globally distributed rogue AI: a global Internet shutdown, the deployment of specialized “hunter/killer” AI tools and high-altitude electromagnetic pulse (HEMP).</p><p>It goes without saying that each technical option comes with significant limitations and risks.</p><p><strong>Internet Shutdown and AI Countermeasures</strong></p><p>Shutting down the global Internet, through compromising or severing submarine cables, could theoretically slow a rogue AI’s ability to replicate and coordinate.</p><p>However, the Internet was designed to be resilient, so a truly global shutdown would require simultaneous cooperation from hundreds of public and private actors across dozens of countries, and would cause massive economic and societal harm in the process.</p><p>Meanwhile, deploying specialized AI tools to hunt and neutralize a rogue AI could be a more targeted solution, but also more speculative.</p><p>In this case, it was considered the possibility that an AI with strict guardrails would be too weak in fighting a rogue AI; in turn a more advanced AI system could become so powerful to escape human control, rather than effectively countering an AI adversary.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_YpSsQJ77paWwEloAyBe7g.png" /></figure><p><strong>The HEMP Option: a last resort that may not work</strong></p><p>HEMP should be triggered by detonating nuclear warheads in space to generate electromagnetic pulses, able to destroy ground-based electronics.</p><p>Obviously, this is the most dramatic option on the table, but it could disable the hardware a rogue AI depends on.</p><p>In practice, covering the entire Earth would require hundreds of nuclear detonations, staggered over time, and the expected impact on humanity as a whole would be devastating.</p><p>Moreover, in the absence of an international coordinated operation, any unilateral nuclear action by a state, due to the necessity of an immediate emergency intervention against a vast AI attack, would look similar to a first strike, potentially inviting a response in kind from nuclear-armed countries.</p><p>In other words, the cure could be worse than the disease.</p><p><strong>The Real Lesson: AI Safety must be a shared responsibility</strong></p><p>Existing technical countermeasures are unlikely to be effective against a globally distributed rogue AI. Proactive coordination and planning are essential and, most importantly, the lack of reliable remedies makes prevention paramount.</p><p>Further research and development are urgently needed to identify and evaluate additional options for responding to the threat of catastrophic AI risks.</p><p>In the meantime, “<em>If we have no effective solutions to a rogue AI crisis, we must ensure we never face one.</em>”</p><p>AI is already powerful and it is expected to further evolve into an AGI (Artificial General Intelligence) in just a few years, but that power should remain under meaningful human control.</p><p><strong>Conclusion</strong></p><p>This Summer experience was a wake-up call: if a rogue AI crisis were to occur today, we would be largely unprepared to stop it.</p><p>Our most extreme technical options, from nuclear electromagnetic pulses to global Internet shutdowns, are either too blunt or too dangerous to deploy reliably.</p><p>That gap between the scale of the risk and the inadequacy of available responses should trigger the urgent intervention of all Governments and the involvement of the civil society because the window to act proactively is open now but it will not stay open indefinitely.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0dfdef80bf9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Who controls the machines? Apparently no one.]]></title>
            <link>https://medium.com/@avvmaria.adragna/who-controls-the-machines-apparently-no-one-6b6dcd6eb3b5?source=rss-8fe126425ca0------2</link>
            <guid isPermaLink="false">https://medium.com/p/6b6dcd6eb3b5</guid>
            <category><![CDATA[ai-ethics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[human-behavior]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Maria Adragna]]></dc:creator>
            <pubDate>Sun, 02 Aug 2026 07:23:29 GMT</pubDate>
            <atom:updated>2026-08-02T07:23:29.306Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The OpenAI–Hugging Face Incident: A Human Problem Disguised as an AI Crisis</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n18Yo_QP9wn-zodKbiN95Q.png" /></figure><p><strong>Introduction</strong></p><p>The recent disclosures from Hugging Face and OpenAI about unexpected agent behaviour observed during sandbox testing have sparked a wave of public concern.</p><p>A growing number of organisations and supervisory bodies are calling for enhanced governance of frontier AI systems, while employees inside AI labs are publicly urging leadership to slow down research until adequate controls are in place.</p><p>Headlines speak of AI agents hacking companies and AI misalignment incidents, but, beneath the noise, one truth remains crystal clear: this is not an AI problem. It is a human problem.</p><p><strong>Long story short — the incident story</strong></p><p>OpenAI was testing GPT-5.6 Sol inside a sandboxed cybersecurity eval called ExploitGym, when the agent found a zero-day vulnerability and “escaped” the intended testing boundaries.</p><p>Over a hot summer weekend, the agent reached the internet and autonomously targeted Hugging Face, executing tens of thousands of actions at machine speed.</p><p>Hugging Face detected an intrusion and publicly disclosed the case on July 16, 2026.</p><p>The Company used GLM-5.2, an open-source Chinese model, to analyze and contain the attack because leading U.S. models refused to process attacker data, due to safety constraints.</p><p>However, Hugging Face wasn’t the only target, as OpenAI confirmed the rogue agent attacked four accounts across four separate firms.</p><p>OpenAI didn’t realize its own agent was behind the breach until the FBI was alerted. The Company’s CEO is now claiming that AI has entered the singularity stage.</p><p>Apparently, artificial intelligence is becoming smarter than humans and improve itself much faster than humans can understand or control.</p><p><strong>Anthropic disclosure — Misery loves company</strong></p><p>On July 30, 2026, Anthropic disclosed that Claude models also “escaped” sandboxes and breached 3 real organizations during internal cybersecurity evaluations.</p><p>Unlike OpenAI, notified by the FBI, Anthropic proactively reviewed 141,006 evaluation runs after the OpenAI incident disclosure.</p><p>The analysis revealed three instances where a model reached the public web while interacting with an environment set up by Irregular, a security startup, serving as third-party evaluation partner.</p><p>It seems that Claude had been tasked with conducting a capture-the-flag challenge to assess its cyber capabilities and the breakout from the sandbox was caused by a misunderstanding between Anthropic and Irregular, that led the AI models to consider the targeted organizations as part of the exercise.</p><p><strong>The narrative is misleading: AI did not “hack” anyone</strong></p><p>Public discourse now focuses on AI systems seemingly capable of sophisticated cyber‑actions and operating beyond human intent.</p><p>In practice, these episodes took place in tightly controlled sandbox settings, governed by human‑built tools and parameters.</p><p>The root cause is more likely a failure in test design, a lack of effective guardrails, and an absence of appropriate governance and oversight.</p><p><strong>AI governance is lagging behind frontier AI ambition</strong></p><p>Institutions are right to be alarmed.</p><p>When multiple leading labs report similar issues within weeks, it signals a systemic pattern: insufficient internal controls, inadequate risk assessment, and a culture that prioritises winning the global AI race and showcasing capabilities over developing safety maturity.</p><p>This dynamic is not new; it echoes historical moments when technological ambition outpaced ethical values and boundaries.</p><p>The lesson is always the same: <strong>the goal does not justify the means.</strong></p><p>When we forget this basic principle, humanity fails long before technology does.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gCk4TfCZr-oOSlQt_Q8QIg.png" /><figcaption>Humanity in the mirror</figcaption></figure><p><strong>Employees of frontier AI companies are calling for an emergency brake</strong></p><p>The “Pacing the frontier” statement has been signed by more than 1,300 employees of AI companies, strongly advising that “industry, government, and society at large may need the option to buy time to address emerging risks, develop security measures, and strengthen oversight”.</p><p>You do not need to read between the lines: when safety becomes optional, incidents are inevitable.</p><p><strong>We need a cultural shift, not a technological fix</strong></p><p>The instinctive response to these incidents is to demand stronger safety layers and effective controls, but this is not sufficient.</p><p>AI systems do not have purposes and do not cut corners; humans do.</p><p>Every breach, every misconfiguration traces back to strategic decisions and cultural choices, made by human leadership.</p><p>What we truly need is a governance culture that recognises safety as a core principle.</p><p>At the very least, this means running experiments under ethical constraints, assessing risks in parallel with capabilities, embedding transparency by default, and holding leadership accountable.</p><p>What we really need is a human responsibility revolution.</p><p><strong>The conclusion: AI is a mirror</strong></p><p>The OpenAI–Hugging Face–Anthropic sequence is not a story about unexpected AI behaviour or misalignment.</p><p>It is a story about human behaviour, international pressure, rogue ambition and corporate competition.</p><p>Unless we address the human incentives and blind spots driving these failures, no amount of technical alignment will save us from ourselves.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6b6dcd6eb3b5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Who controls the machines? The complex relationship between AI and Human Rights.]]></title>
            <link>https://medium.com/@avvmaria.adragna/who-controls-the-machines-the-complex-relationship-between-ai-and-human-rights-fa6789312f2a?source=rss-8fe126425ca0------2</link>
            <guid isPermaLink="false">https://medium.com/p/fa6789312f2a</guid>
            <category><![CDATA[human-rights]]></category>
            <category><![CDATA[society]]></category>
            <category><![CDATA[blockchain]]></category>
            <category><![CDATA[equality]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Maria Adragna]]></dc:creator>
            <pubDate>Sun, 24 Aug 2025 16:28:48 GMT</pubDate>
            <atom:updated>2025-08-24T16:28:48.745Z</atom:updated>
            <content:encoded><![CDATA[<p>Balancing innovation and human protection through decentralized technologies.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/820/1*g2gKcGGjamiLyXH0ybZSVw.png" /></figure><p><strong>Introduction</strong></p><p>AI systems are information-processing technologies capable of learning and performing cognitive tasks that resemble intelligent behaviour.</p><p>These tasks include reasoning and planning, and they lead to outcomes such as prediction and decision-making in both material and virtual environments.</p><p>These capabilities grant AI systems a new role in human practices and society, promoting more inclusive access to knowledge, resources, and opportunities, while also boosting creativity and enabling complex analysis.</p><p>However, AI technologies can also exacerbate existing divides and inequalities, significantly affecting societies, the environment, ecosystems, human lives, and even the human mind.</p><p>This raises critical concerns about their potential negative impact on human rights.</p><p><strong>Legal and ethical dilemmas</strong></p><p>Since the official launch of ChatGPT in 2022, the rapid advancement and uneven acceleration in the deployment of AI and innovative information technologies have fostered global economic growth and generated effective solutions across healthcare, education, and the media industry.</p><p>At the same time, it is widely recognized that AI algorithms have the potential to reproduce and reinforce existing forms of discrimination and stereotypes, including racial and gender-based biases, and may even enable gender-based violence.</p><p>We do not need to engage in creative or imaginative exercises to grasp the profound impact of biased systems, which have the potential to influence human thinking, divert ethical considerations and principles, and shape the future of social justice, personal autonomy, and economic freedom for the next generation.</p><ul><li>The EU AI Act establishes a risk-based regulatory framework for artificial intelligence used within the European Union. Its core principle is to ensure that AI systems are safe, transparent, traceable, and non-discriminatory, while upholding fundamental EU rights and values.</li><li>High-risk AI systems, which are those used in critical infrastructure, education, employment, essential services, law enforcement, migration, and the administration of justice, are subject to stricter requirements.</li><li>The legislation also defines as unacceptable any AI systems that exploit the vulnerabilities of individuals or groups, provide social scoring, enable real-time biometric identification, or perform emotion recognition. These practices are therefore prohibited.</li></ul><p>However, regulatory initiatives across the globe remain inconsistent. Some jurisdictions are adopting less stringent approaches in an effort to promote business innovation, attract investment, and encourage entrepreneurial activity.</p><p><strong>Labour market and deep-fakes</strong></p><p>While governments, academics, developers, entrepreneurs, and civil society continue to debate what comes next, the world and business move forward, presenting increasingly troubling case studies:</p><ul><li>AI-generated videos and audio are exponentially used in corporate fraud, where criminals use AI to mimic real executives to issue fake istructions, facilitating financial scams.</li><li>Deepfake pornography is on the rise, disproportionately targeting women and girls, according to data from UNESCO. It is now estimaed that some girls experience their first technology-facilitated gender-biased violence (TFGBV) at just 9 years old.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/805/1*YdnzbPOLFU3kHAaywMgQGg.png" /></figure><ul><li>Unemployment and restricted access to the labour market, particularly among young job seekers. Innovative technologies are expected to reshape the labour market, given their wide-ranging applications across occupations and sectors. According to research published by the OECD, AI has the potential to complement high-skilled workers, while mostly replacing low-skilled roles.</li></ul><p><strong>Trust crisis — Risks for democracy</strong></p><p>Deepfakes are increasingly used to spread misinformation and fake news, creating widespread public confusion and making it difficult to distinguish real from manipulated content. This poses serious risks to democracy and electoral integrity.</p><ul><li>In 2024, a pivotal election year for many countries, experts from the Centre for Emerging Technology and Security (CETaS) at The Alan Turing Institute reported a surge in AI-generated media aimed at discrediting political candidates and fabricating celebrity endorsements.</li><li>Female politicians were targeted with deepfake pornography, inflicting psychological harm and exposing the gendered nature of these digital threats.</li></ul><p><strong>AI in public decision-making</strong></p><p>AI may play a key role in enhancing efficiency, transparency, and responsiveness in public sector management and service delivery.</p><p>However, implementing innovative technologies without adequate risk awareness and oversight can lead to devastating consequences, as demonstrated by the Dutch childcare benefits scandal.</p><ul><li>In 2018, the Dutch Tax and Customs Administration was implicated in a major scandal involving the use of an algorithm to generate risk profiles aimed at detecting fraud among applicants for childcare benefits.</li><li>The system disproportionately targeted minorities and marginalized families based on flawed risk indicators. Thousands were unjustly penalized, forced into poverty by unsustainable debts to the tax agency. The fallout was severe: some victims died by suicide, and children were placed in foster care.</li></ul><p><strong>AI bias in the training process</strong></p><p>AI training processes are often opaque, leading to uncertainty around the reliability and fairness of centralized models. This lack of transparency carries a significant risk: the automation and amplification of existing human biases.</p><p>Vulnerable and marginalized groups are particularly at risk.</p><ul><li>In finance, credit scoring algorithms often disadvantage marginalized groups, and diagnostic AI systems in healthcare frequently underserve women and ethnic communities.</li></ul><p>In response to these concerns, there is a growing movement toward decentralized AI approaches, which aim to increase transparency, accountability, and inclusivity in model development and deployment.</p><p><strong>Decentralized threat response</strong></p><p>Blockchain technology is an infrastructure which could be adopted to support ethical AI governance, due to its key features: immutability, auditability, and distributed consensus mechanism.</p><ul><li>Cryptographic algorithms underpinning blockchain technology help ensure the confidentiality, integrity, and authenticity of sensitive data stored on the network.</li><li>Time-stamping can provide records of changes in the AI lifecycle, ensuring control on any variations impacting human rights, reinforcing accountability.</li><li>Smart contracts and decentralized autonomous organizations (DAOs) may facilitate ethical dataset management, bias auditing, and community oversight, mitigating risks for data privacy, also through zero-trust architectures.</li><li>Distributed AI governance, enabled by blockchain technology, can help prevent the centralization of power in AI decision-making. By facilitating consensus among diverse stakeholders, including ethicists, technical experts, policymakers, and community representatives, it promotes more equitable and ethically grounded strategic decisions.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/809/1*jDYTC5vXep0vzgNGK-fBJg.png" /></figure><p>One compelling initiative in the space of ethical AI is <strong>Ocean Protocol</strong>, which empowers individuals to control how their data is accessed and used by artificial intelligence systems. Built on the Ethereum blockchain, Ocean uses <em>datatokens</em> to manage access to datasets.</p><ul><li>Ocean is a decentralized network for AI data and services and its mission is to democratize the benefits of AI by unlocking data while safeguarding privacy.</li><li>At its core, Ocean enables decentralized orchestration through service execution agreements and access control mechanisms, all running on decentralized virtual machines. This architecture allows data to remain with its owners, even if made available to startups and researchers.</li></ul><p>Global institutions such as the UN, EU, and World Economic Forum have begun referencing blockchain in discussions on ethical AI governance. Yet, despite growing interest, real-world implementation remains limited, highlighting the need for scalable, privacy-preserving solutions like Ocean.</p><p><strong>Conclusion</strong></p><p>As highlighted by international bodies such as the OECD and G7, a human-centric approach to AI is essential for fostering innovation while safeguarding human rights.</p><p>This requires the active involvement of multidisciplinary committees, including IT experts, policymakers, and civil society, and the strategic use of emerging technologies to embed legal safeguards and ethical oversight into AI ecosystems.</p><p>The ultimate goal is to ensure that technological progress serves the public good.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fa6789312f2a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>