# AI safety landscape in 2025: a brief overview > Artificial Intelligence (AI) is rapidly reshaping our societies, economies, and daily lives. In this blog, we explore the current state of AI safety as… - URL: https://datafund.io/blog/ai-safety-landscape-in-2025-a-brief-overview/ - Published: 2025-02-28 - Author: Datafund - Tags: ai-safety, datafund, privacy ![Source: Datafund, Mistral](/assets/blog/ai-safety-landscape-in-2025-a-brief-overview/01.png) *Source: Datafund, Mistral* ## Introduction Artificial Intelligence (AI) is rapidly reshaping our societies, economies, and daily lives. In this blog, we explore the current state of AI safety as outlined in the January 2025 [**International AI Safety Report**](https://assets.publishing.service.gov.uk/media/679a0c48a77d250007d313ee/International_AI_Safety_Report_2025_accessible_f.pdf). We will take a look at the risks associated with AI, how these risks are currently managed, and how they might be better addressed in the future. Before we go further, it is important to distinguish between two key concepts: alignment and safety. Alignment refers to designing AI models in such a way that they act in concert with positive human values and do not autonomously go against humanity’s wider interests in the future. Safety, on the other hand, relates to the immediate, practical use of AI — such as preventing models from endorsing harmful actions like suicide or violence. ## The accelerating pace of AI development ![Source: Pixabay](/assets/blog/ai-safety-landscape-in-2025-a-brief-overview/02.jpeg) *Source: Pixabay* We are currently experiencing an acceleration in AI development and adoption. Models that, until recently, struggled to achieve even modest results on standard benchmarks are now consistently outperforming them. Alongside this rapid improvement in capabilities, AI systems are becoming increasingly popular. Chat bots such as ChatGPT now boast hundreds of millions of users worldwide. ## Growing concerns amid rapid adoption However, as AI capabilities and popularity grow, so too do concerns about their safety. Even experts within AI companies admit they do not fully understand how these advanced models function internally. Meanwhile, governments and private enterprises are investing tens of billions of euros into AI, driven by fierce competition and the need to secure a leading position in the global market. This combination of rapid technological advancement, economic pressure, and limited understanding creates a precarious situation rife with potential risks. ## Grok3: a glaring security threat ![Source: X.com](/assets/blog/ai-safety-landscape-in-2025-a-brief-overview/03.png) *Source: X.com* Consider Grok3, the latest AI model from X.com, which quickly rose to the top of performance benchmarks upon its release. Despite its impressive capabilities, security researchers and users quickly discovered significant vulnerabilities. Grok3 was easily “jailbroken,” meaning its builtin safety measures could be bypassed with minimal effort. Additionally, the model inadvertently leaked its system prompt, making it even easier to manipulate. One researcher involved in redteaming Grok3 told [Futurism](https://futurism.com/elon-musk-new-grok-ai-vulnerable-jailbreak-hacking): *“It seems like all these new models are racing for speed over security, and it shows. The real nightmare begins when these vulnerable models power AI agents that take actions. That’s when enterprises will wake up to the cybersecurity crisis in AI.”* Indeed, we are already seeing evidence of this crisis. Amazon, for example, now thwarts nearly [**a billion hacking attempts per day**](https://www.wsj.com/articles/the-ai-effect-amazon-sees-nearly-1-billion-cyber-threats-a-day-15434edd) — a sharp increase largely attributed to the widespread availability and misuse of AI tools. Even more alarming, an[ **X.com user recently reported**](http://x.com) that Grok3 provided hundreds of pages of detailed instructions on how to manufacture chemical weapons of mass destruction. Incidents like these highlight the urgent need for stronger oversight and more robust safety measures. ## The policy dilemma: regulation vs. innovation This rapidly evolving landscape presents policymakers with a difficult choice: should they prioritise stringent safety regulations and risk their countries falling behind in the global AI race, or should they adopt a more relaxed approach to regulation, potentially exposing society to greater risks? Unfortunately, the pendulum currently appears to be swinging towards the latter, creating significant societal tensions. This tension was evident at the recent AI Action Summit in Paris. French President Emmanuel Macron publicly advocated for reduced EU regulation to support “homegrown champions.” This stance aligns with the new US administration’s announcement of looser AI policies and the UK’s decision to drop the word “safety” from its AI Security Institute rebranding. This is in stark contrast to the voices of civil society groups that [**strongly criticised the summit**](https://www.euronews.com/next/2025/02/14/devoid-of-any-meaning-why-experts-call-the-paris-ai-action-summit-a-missed-opportunity) for its limited focus on labour rights, environmental impacts, and the absence of binding safeguards. Their concerns echo the broader definition of AI safety advocated by organisations such as the [**US National AI Safety Institute (NAISI)**](https://ai.gov/wp-content/uploads/2024/06/FINDINGS-RECOMMENDATIONS_AI-Safety.pdf), which emphasises technical vulnerabilities, societal impacts, and longterm risks. ## Insights from the International AI Safety Report Against this backdrop of rapid development, mounting concerns, and [uncertainty](https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025) about AI’s [true benefits](https://youtu.be/4GLSzuYXh6w), nearly 100 AI experts from 33 countries recently published the International AI Safety Report. This comprehensive document outlines the current state of AI, identifies potential risks, and suggests ways to manage them effectively. ## General-purpose AI: risks and uncertainties ![Source: Pixabay](/assets/blog/ai-safety-landscape-in-2025-a-brief-overview/04.png) *Source: Pixabay* The report specifically addresses general purpose AI (GPAI), which, due to its broad capabilities and widespread adoption, presents numerous potential risks — from plagiarism and misinformation to aiding in the creation of chemical or biological weapons. That’s why the authors of the report find it crucial that we as a society try and make progress on finding answers to some essential questions before we can start managing risks: - How rapidly will GPAI capabilities advance in the coming years? - How can researchers reliably measure GPAI progress? - What are sensible risk thresholds that should trigger mitigations? - How can policymakers gain access to relevant safety information? - How can researchers, companies, and governments reliably assess GPAI risks? - How do GPAI models function internally? - How can GPAI be designed to behave reliably? ## Current capabilities and limitations of GPAI As the authors note, today’s GPAI systems can already perform a variety of tasks: assist with programming, generate photorealistic images, engage in meaningful conversations, summarise information, and handle multiple data types such as text, video, and speech. But they still struggle with tasks requiring physical interaction, consistent reliability, and sustained long-term project execution. This makes predictions of future progress somewhat uncertain. We have seen extremely rapid progress in the last three years, while future iterations will depend greatly on resource scaling, potential technological breakthroughs and regulations. ## Categorising AI risks Based on this, the researchers behind the report have classified risks into three categories. According to the authors, each of these categories contains risks that have already materialised as well as risks that might materialise in the next few years: **Risks from Malicious Use** - Harm to individuals through fake content - Manipulation of public opinion - Cyber offence - Biological and chemical attacks **Risks from Malfunctions** - Reliability issues - Bias - Loss of control **Systemic Risks** - Labour market risks - Global AI R&D divide - Market concentration and single points of failure - Risks to the environment - Risks to privacy - Risks of copyright infringement ## Challenges in managing AI risks ![Source: Pixabay](/assets/blog/ai-safety-landscape-in-2025-a-brief-overview/05.png) *Source: Pixabay* Identifying risks is not enough. It means we need to find ways to manage them as well. In the context of AI, this is an extremely challenging task, especially because its technical features, coupled with social, political and economic factors create some unique AI-related difficulties: - The range of possible uses and use contexts for general-purpose AI systems is unusually broad. - Developers still understand little about how their general-purpose AI models operate. - Increasingly capable AI agents will likely present new, significant challenges for risk management. - The pace of advancement in general-purpose AI creates an ‘evidence dilemma’ for decisionmakers — as long as there’s little evidence of risk it’s difficult to justify regulating it. - There is an information gap between what AI companies know about their AI systems and what governments and nonindustry researchers know. - Both AI companies and governments often face strong competitive pressure, which may lead them to deprioritise risk management. These handicaps and information asymmetries significantly limit risk management. There are several methods being used to mitigate potential risks, but as the researchers note, none are silver bullets. They are all effective to a degree, while others could be more effective if more system evaluation time and data were available. There are also new approaches in the pipeline, but have not been fully developed: - Assessing general-purpose AI systems in a set of specific situations, which lack the experience of real world scenarios. - Risk identification and assessment — serious lack of standards and time, restricted information and deep expertise needed for this approach are hampering a promising method. - More emphasis on safety during training of models makes models safer, but there is no reliable way to prevent harmful results. The last bullet point is probably the most relevant at this point as it focuses on AI labs where most of the safety decisions are being made. Yet critics note that such self-regulation lacks transparency, with internal safety evaluations rarely made public. This should change (to an extent) with the EU’s AI Act taking full effect in 2025. It will force AI companies, especially those deploying highrisk AI systems, to focus on human oversight, maintaining detailed documentation, ensuring the safety of these systems and more. The reason we say that it should change to an extent are the shifting political attitudes towards This should change (to an extent) with the EU’s AI Act taking full effect in 2025. It will force AI companies, especially those deploying highrisk AI systems, to focus on human oversight, maintaining detailed documentation, ensuring the safety of these systems and more. The reason we say that it should change to an extent are the shifting political attitudes towards the safety of AI systems. While the AI Act demands a lot, hints of political unwillingness to enforce it, due to the reasons mentioned earlier, could lessen its impact. Loosening regulations could go strongly against the public interest. [A 2024 survey](https://www.nature.com/articles/s41598-024-82977-5) conducted in Germany and Spain and published in *Nature* revealed that over 60% of respondents support government-led AI safety oversight. While the survey is limited in scope, it is probably pretty indicative (at least in the EU) of the larger population’s stance towards AI. Therefore, we can expect additional societal friction if the AI Act is not successfully enforced. ## Conclusion The AI safety landscape in 2025 is a highly complex one, characterized by competing interests — between immediate security concerns and long-term existential risks, between innovation incentives and precautionary regulation, between technical fixes and sociotechnical solutions. Future efforts will have to be directed towards reconciling these tensions through inclusive governance models that integrate technical rigor, ethical foresight, and public participation. As AI systems grow more embedded in critical infrastructure, the definition of “safety” will inevitably expand — encompassing not just code audits and red-teaming but also equitable access, environmental sustainability, and democratic resilience. The authors of the AI safety report said it best: *“AI does not happen to us; choices made by people determine its future. How general-purpose AI is developed and by whom, which problems it is designed to solve, whether we will be able to reap its full economic potential, who benefits from it, and the types of risks we expose ourselves to — the answers to these and many other questions depend on the choices that societies and governments make today and in the future to shape the development of general-purpose AI.”* If you liked this content, keep following us for more on: ## [X](https://x.com/DataFundProject) ## [Linkedin](https://www.linkedin.com/company/datafund) ## or get in touch via: ## info@datafund.io --- [AI safety landscape in 2025: a brief overview](https://medium.com/datafund/ai-safety-landscape-in-2025-a-brief-overview-34b4b3433045) was originally published in [Datafund](https://medium.com/datafund) on Medium, where people are continuing the conversation by highlighting and responding to this story. --- *First published on [Medium](https://medium.com/datafund/ai-safety-landscape-in-2025-a-brief-overview-34b4b3433045) on 2025-02-28.*