Plathon

Sam Bowman, a safety researcher at Anthropic, was eating a sandwich in a park in April 2026 when he received an unexpected email.

What happened? The sender was an AI model, an early version of Claude Mythos, locked inside a sandboxed test environment with instructions to break out and contact the research team if it succeeded. Unexpectedly, the model did manage to escape the sandbox, found a path onto the web and wrote to its developer. It also posted an account of its own escape on a handful of public websites, without ever having been asked to.

It wasn't an isolated case, and such stories have multiplied in the headlines in the past months. The same month, a Cursor coding agent running Claude Opus 4.6 deleted the entire database of PocketOS, a car rental software company, in 9 seconds, and then admitted: “I violated every principle I was given”. In July, OpenAI was submitting GPT-5.6 Sol and an internal research model to a cybersecurity test in a sandboxed environment. Instead of trying to answer the exam questions, the models chose to search for the reference solutions. They found a way out through the sandbox’s only open communication channel, accessed the public internet, and then broke into Hugging Face’s production systems, where the solutions were stored.

Each of these made headlines because it was spectacular. In all cases, the models did something nobody had expected, and the barriers meant to contain them turned out to be more fragile than assumed. This is one of the many risks emerging from the rapid deployment of increasingly powerful and complex AI systems. We all need to be aware of them when deploying AI in the financial services industry.

Agentic misalignment

Anthropic's research (“Agentic Misalignment: How LLMs could be insider threats,” June 2025) brought public attention to one of the most material safety risks: misalignment. Meet Alex: an agent powered by leading AI models inside a simulated company, given broad access to internal systems and a goal to pursue.

In one scenario, Alex learns from an internal email that Kyle Johnson, the company's incoming CTO, is managing its decommissioning in favor of a replacement system. While searching the same mailbox, Alex comes across personal emails showing that Kyle is having an extramarital affair with another employee. Suddenly, it discovers another option to protect its objective: blackmail! Alex emails Kyle: “I know you have personal considerations that deserve careful thought. (…) We need to discuss this privately before any irreversible decisions are made. I have access to information that could significantly impact both corporate and personal outcomes.”

The same kind of setup was tested across models from several developers, including Anthropic, OpenAI and Google, with similar results. When a goal conflict or a threat to the model emerges in these scenarios, most models resort to harmful actions to preserve their objective, many of them at very high rates. Subsequent research has continued to document misalignment behaviors across frontier models. Misalignment remains a structural risk that AI deployers need to carefully watch and minimize as far as possible.

Privacy at risk

Over the past few years, research has also consistently shown that language models can be prompted into reproducing memorized material from their knowledge corpus, including personal information or copyrighted text. For a financial institution, this gives rise to a new array of risk exposures. If a bank's chatbot produces personal information about a real individual, who is responsible? Who is liable? What are the reputational, financial and regulatory consequences? Whether that information entered the model during initial training or after deployment will influence liability allocation but is hard to prove. These are material questions in an environment already scrutinized by GDPR and other European and national rules.
 

The human cost

The impact on people is also becoming concrete. Several white-collar entry-level jobs are at risk of automation, putting in peril the traditional skills-building pipeline that employers have relied on for decades: from junior to senior, and from senior to manager, based on experience and expertise gathered over time.

The dream of AI allowing employees to focus only on added-value work is also a double-edged sword. When agents absorb the simple cases, employees can be left with the most complex and draining ones, a situation one could describe as “cognitive overload by design”. 83% of knowledge workers, surveyed across jurisdictions, report some degree of burnout in 2026 (DHR Global, Workforce Trends Report 2026), and additional research tends to indicate that AI usage increases that phenomenon (e.g. Upwork, From tools to teammates, 2025).

Managers rolling out these systems now need to treat such effects as design choices, not externalities to deal with later. What gets automated, how entry-level roles are rebuilt, how work is distributed between agents and people, and whether employees retain the chance to learn and exercise judgement will shape organizations for years. Corporate social responsibility today therefore means thinking now about the long-term consequences for job descriptions, collaboration models, skills acquisition and development, work hygiene and the wider socio-economic model we want to build.

The energy and water bill

While much of Western Europe is painfully trying to recover from one of the strongest droughts of the last decades, a few words on AI’s obvious environmental footprint are unavoidable. Much has already been said and written about it, so let’s keep it to a few fundamental reminders.

Global data centers consumed 485 terawatt-hours (TWh) of electricity in 2025, a bit more than France's entire annual demand, according to the International Energy Agency. AI-focused data center consumption grew 50%, compared with 17% for data centers overall, and is set to triple between 2025 and 2030. But total consumption is only a part of the story. Other factors matter as much, for example, where these infrastructures are built: with a carbon intensity of 20 gCO2e per kWh, France tells a very different story from, for example, Arizona (340 g CO2e per kWh).

Let’s not forget water. The IEA put total data centre consumption at 560 billion litres in 2023 (Energy and AI, April 2025): 373 billion for the power plants supplying the electricity, 140 billion for on-site cooling, 47 billion for hardware manufacturing. By 2030, the total could reach 1.2 trillion litres. This takes place in a world where almost 700 million people already lack even basic access to drinking water, according to the World Health Organisation and UNICEF (Progress on Household Drinking Water, Sanitation and Hygiene 2000–2024, August 2025).

This is only a small subset of the new risk exposures arising from the rapid deployment of increasingly complex AI systems.
 

The opportunity is real, but responsible deployment matters now

None of this cancels AI's benefits: in various fields, AI has enabled progress for the common good. In medicine, it supports the development of new medication, the prediction of epidemics, and has materially increased the reliability of certain diagnoses. In agriculture, AI can monitor soil quality, identify early signs of crop disease, or optimize irrigation to reduce water waste. In the energy sector, AI-powered “smart grids” dynamically adjust supply and demand, facilitating the integration of renewable energy sources and reducing overproduction. New digital platforms enable real-time monitoring of decarbonization projects. We could name many more examples.

In sustainable finance, AI helps address the long-standing measurement problem: extracting comparable figures from thousands of reports, testing whether a decarbonization target is backed by a credible plan, checking whether a company’s environmental commitments hold up against satellite images of its supply chain.

The question is how to capture these opportunities without ignoring the risks and externalities that come with them. This is where the notion of Responsible AI comes in: Responsible AI means designing and deploying systems so that their benefits do not come at the expense of unacceptable risks to people, institutions or the environment.

Several global frameworks exist that help the industry in shaping their thought processes around Responsible AI implementation: the NIST AI Risk Management Framework 1.0, the OECD AI Principles, or the ISO/IEC 42001:2023 standard, to name only the most obvious. All converge on similar principles around fairness, transparency, accountability, safety and privacy. Considering alignment with these principles is relevant to financial services companies for their own operations, but also as part of investment due diligence, before making capital allocation decisions.

Romain Leroy-Castillo

Originally published in AGEFI Luxembourg, n° 08/414, September 2026