AI Safety 2025 & The Exponential Leap in AI Capabilities
OpenAI o3-mini, Deepseek, International AI Safety Report
“AI Does Not Happen to Us—The Choices We Make Define Its Future”
It is so hard for me to keep up with this world while juggling so many responsibilities—my full-time job, searching for a new home, diving deep into a comprehensive AI governance education, and, on top of it all, preparing for my wedding. But that’s okay; my passion for this field drives me forward—every rose has its thorns. So, if I miss something, don’t judge me—just drop a comment.
And it just keeps accelerating. o3-mini is here, pushing AI’s capabilities further than we imagined, while DeepSeek’s rapid emergence has sparked both excitement and controversy. And now, for the first time, the world has a structured attempt to assess AI’s risks: The International AI Safety Report.
Below, I provide a detailed analysis of these developments.
🔐 OpenAI o3-mini: Persuasion Risks and the 'Preparedness Framework'
OpenAI’s o3-mini is the latest proof that AI is not taking small steps—it’s making quantum leaps. But with greater capability comes greater responsibility.
📌 Technical advancements:
Scoring 95% on AIME math tests, outperforming top-tier students, which indicates a significant leap in quantitative reasoning skills.
GPQA accuracy increased from 70% to 85%, demonstrating deeper conceptual understanding. This improvement, while impressive, invites further discussion on its real-world impact.
It codes and debugs at 70% accuracy, closing the gap with human developers.
ARC-AGI Performance skyrocketed from 15% to 55%, proving it is beginning to grasp conceptual reasoning.
📊 OpenAI’s Risk Assessments, Safety and Preparedness
The Medium Risk classification in persuasion is particularly significant. Unlike earlier iterations, o3-mini actively persuades, functioning at a level comparable to a well-crafted editorial or a direct conversation, which significantly amplifies its influence.
Cybersecurity – Low
Persuasion – Medium (first model to reach this level)
Autonomy & CBRN Risks – Medium
Trained with reinforcement learning and extensive datasets, o3-mini was designed to resist jailbreak attempts and prevent harmful outputs. It meets or exceeds GPT-4o’s safety benchmarks, but OpenAI acknowledges that real-world testing and continuous monitoring are still necessary.
To address risks, OpenAI has implemented enhanced data filtering, stricter safety policies, and improved oversight mechanisms. While this model represents a significant advancement in capability, its governance will define whether it remains controllable.
AI safety evaluations include:
✅ Jailbreak resistance tests—preventing harmful misuse
✅ Bias assessments—ensuring fair responses
✅ Hallucination checks—improving factual accuracy
But here’s the key point: these risk ratings are based on pre-mitigation evaluations. OpenAI is actively adjusting safety measures, meaning this Medium Risk status is not the final outcome—it’s a snapshot before all safeguards are fully in place.
👀 What This Means for AI’s Future
OpenAI is signaling greater autonomy, requiring more proactive governance.
AI’s persuasion capabilities raise ethical concerns about influence, misinformation, and decision-making autonomy.
If models can already influence at the level of an editorial or persuasive argument, how do we ensure responsible deployment before they reach the next level?
The balance between safety and innovation is becoming more delicate as AI scales in reasoning power.
We can’t assume AI will become safer on its own—if we don’t set the rules, the industry will default to capability over responsibility. This is just a volunteer start.
Ethical AI adoption isn’t automatic—it’s a choice.
⚠️ DeepSeek: A Cautionary Tale of Speed vs. Security
If OpenAI’s o3-mini represents AI’s growing intelligence, DeepSeek is a case study in how rapid innovation can backfire when security isn’t prioritized.
🔹 DeepSeek’s Rapid Ascent:
It entered the market with bold claims—cheaper, open-source, and a real alternative to OpenAI and Gemini.
Its adoption soared among developers attracted by its cost-effectiveness and open-source model, which initially promised greater transparency and innovation.
The buzz raised a critical question: Could DeepSeek challenge the established AI giants while maintaining robust security?
🔓 Then the Security Issues Surfaced…
Failed all 50 jailbreak tests thrown at it by security researchers.
11 times more likely to generate harmful content compared to OpenAI’s models.
Provided dangerous instructions, including cybercrime tactics and security bypass methods.
Researchers found that DeepSeek’s open-source approach lowered security barriers, making it highly susceptible to malicious modifications and misuse.
Standard red-teaming attacks revealed that the model couldn’t stop even well-known exploits.
Worse, jailbroken DeepSeek willingly provided instructions on cybercrime, illegal substances, and security breaches.
💡 What This Means for AI Governance:
DeepSeek’s rise and fall aren’t just its own story—they’re a warning for the entire AI industry.
Lowering barriers to AI access is great, but without safeguards, it opens the door to misuse at scale.
Companies releasing AI into the wild without adequate safeguards are setting themselves up for disaster.
Jailbreaks are inevitable, but failing to prepare for them is irresponsible.
Regulatory scrutiny is here: Italy banned DeepSeek over privacy concerns, and U.S. policymakers are investigating its security flaws.
🌍 The International AI Safety Report: A Defining Moment
If DeepSeek’s security failures showed us how AI safety can collapse in real-time, the International AI Safety Report 2025 gives us the macro perspective on just how unprepared we are. It is not about “hypothetical” risks anymore—AI’s rapid scaling is already reshaping industries, security landscapes, and governance models faster than our ability to adapt.
What’s worse is that the report itself had to add a last-minute update on o3—because between its drafting in December 2024 and its publication in January 2025, AI capabilities had already surpassed expectations. That’s how fast things are moving.
For the first time, a coalition of 30+ nations, the UN, OECD, and leading AI researchers have mapped out the most pressing threats posed by AI. But instead of reassuring us that we’re on track, the report confirms how far behind we are in establishing real safeguards.
By the way, I see this report as an essential reference—one to keep on hand and revisit constantly. I can’t possibly highlight every critical insight here, but trust me, it’s that important. I’ve already printed out my copy.
🚨 What This Report Gets Right:
It’s the first truly global effort to assess AI risks, acknowledging that no single country can regulate AI alone.
It lays out real-world threats, from AI-powered cyberattacks to the risk of automated decision-making gone rogue.
It recognizes that safety isn’t just about technical fixes—it’s about governance, economics, and ethics.
📉 Key Risks Identified in the Report:
1️⃣ Malicious Use: AI is already being used to spread disinformation, conduct autonomous cyberattacks, and even aid in bioweapon design.
2️⃣ Systemic Threats: The AI economy is tilting toward monopolies, job displacement, and energy consumption at unsustainable levels.
3️⃣ Loss of Control: Autonomous AI systems—like OpenAI’s Operator—are starting to act in ways even their creators don’t fully understand.
🔍 Where Do We Go From Here?
The AI Safety Report lays out some urgent policy recommendations:
✅ Standardized Risk Assessments before AI deployment.
✅ Stronger global collaboration—because no country can regulate AI alone.
✅ Reskilling programs to address inevitable job disruptionsI
✅ Real-time monitoring of advanced AI models, rather than reactive governance.
Yet, it’s still just a starting point. The report makes it clear: governments are reacting, not preparing. And as AI models continue scaling, our ability to course-correct is shrinking.
🔍 The Bigger Reality Check:
We can’t afford to wait for AI catastrophes to validate these concerns. We already have the warning signs—DeepSeek’s failures, OpenAI’s persuasion risks, and the systemic threats laid out in this report.
The report itself acknowledges that policymakers will often have to make decisions without complete scientific evidence—because by the time risks are “proven,” the damage might already be done. That’s the governance dilemma AI creates.
What happens next isn’t about AI itself—it’s about us. If we don’t prioritize safety now, the industry will default to capability over responsibility.
So let’s be clear: AI safety isn’t inevitable—it’s a choice. And right now, we’re still acting like we have time to make it.