Welcome back! The temperature around AI safety just spiked. Safety warnings are turning into global coordination calls. Corporate adoption is widespread but uneven, with many still struggling to prove the payoff. Security researchers are facing systems that mutate faster than they can defend, and scientists are mapping how artificial minds actually store what they know.
In today’s Generative AI Newsletter:
• OpenAI warns of “catastrophic misuse” as intelligence accelerates
• McKinsey finds most firms still failing to turn AI pilots into progress
• Google detects malware using LLMs to evolve mid-attack
• Researchers uncover separate logic and memory circuits inside AI models
Latest Developments
OpenAI has issued a new call for international coordination on AI safety, saying the world may be only a few years away from AI systems that can make scientific discoveries. The company believes today’s models already outperform top humans on complex reasoning tasks and are “about 80% of the way to an AI researcher.”
What OpenAI expects next:
AI Milestones: Small discoveries expected by 2026, major breakthroughs by 2028, with intelligence costs dropping 40x per year.
Safety Collaboration: Governments and labs urged to coordinate early on oversight and technical safeguards.
Shared Standards: Top labs encouraged to align on safety protocols and publish evaluations to avoid unsafe competition.
Resilience Systems: OpenAI recommends building an AI safety infrastructure modeled on cybersecurity.
Public Accountability: Regular reporting and global monitoring of AI’s real-world effects to guide policy.
OpenAI’s tone was part warning and part plea. The company seems less worried about when superintelligence will arrive and more about whether humanity will be ready when it does. Intelligence is multiplying faster than wisdom, and that imbalance could shape the century.
Special highlight from our network
Tired of writing meeting follow-ups? Same
You know the feeling: the meeting ends, and suddenly you’ve got homework.
Notes to tidy, action items to assign, emails to draft.
The new AI personal assistant app Radiant handles all that for you. And did we mention it’s totally free while it’s still in beta?
With Radiant all your summaries, next steps, and action items land in one neat email draft, ready to review and send in seconds.
McKinsey’s new State of AI 2025 report paints a picture of near-universal AI adoption with surprisingly limited results. Out of nearly 2,000 organizations surveyed, 88 percent say they now use AI somewhere in their business, yet only one-third have scaled it beyond pilot projects.
Here’s what McKinsey found:
Widespread but shallow use. Most companies remain stuck in pilot mode, testing ideas rather than transforming operations.
Modest business impact. Only 39 percent reported profit gains from AI, and just 6 percent saw earnings rise more than five percent.
Early agent adoption. About 62 percent are exploring AI agents, though only 23 percent have scaled them, mostly for IT and knowledge management.
Workforce shifts ahead. Roughly 32 percent of firms expect job cuts of three percent or more next year, while only 13 percent foresee hiring growth.
High performers break the mold. The few companies seeing real value use AI to redesign workflows, link departments, and drive innovation rather than just efficiency.
McKinsey’s takeaway is blunt. The AI revolution has reached almost every company, but only a handful know what to do with it. The winners are treating AI less like a gadget and more like a new management philosophy, turning experimentation into strategy while everyone else runs pilots that never take off.
Google’s threat team has spotted a worrying shift. Malware is moving beyond being written with AI assistance. Some samples now call large language models while they run to rewrite themselves, hide their behavior, and dodge detection. That makes these strains more adaptive and harder to pin down, which changes the rules for defenders and raises the stakes for model safety.
Here’s what to know
Promptflux example Promptflux prompts an LLM mid-execution to produce obfuscated code and then saves the new version to persist on a machine.
Other families Variants named Fruitshell, Promptlock, Promptsteal, and Quietvault use LLMs to generate scripts, mine data, steal credentials, or evade LLM-aware scanners.
How it works The malware sends context to an LLM, receives code or commands, then executes that output immediately. That gives attackers a dynamic toolkit instead of a static payload.
Why defenders struggle Traditional signatures fail when code rewrites itself. Detection now requires behavioral tracing and provenance checks across the whole execution chain.
Google’s response Google says it used lessons from these samples to harden both model behaviour and its detection classifiers so models refuse harmful requests more reliably.
This is AI abuse with moving parts. Attackers have stolen a page from living systems by making their tools adapt on the fly. The good news is defenders can borrow a page back by adding real-time tracing, provenance signals, and model hardening. The bad news is that the next wave of attacks will be cleverer, faster, and slipperier.
A new study from has mapped what might be the clearest boundary yet inside a neural network’s mind. The team found that AI models store memorization and reasoning in completely separate neural pathways, almost like two different processors working side by side. When researchers removed the “memorization” circuits, the model lost nearly all recall ability but kept its logic intact.
What they discovered:
Clean split inside models. Removing memory circuits erased 97 percent of the model’s ability to repeat training data, while reasoning tasks stayed steady.
Arithmetic lives with memory. Math performance collapsed by a third when memory circuits were cut, showing that basic calculation behaves more like recall than reasoning.
Curvature mapping. Using a method called K-FAC, researchers visualized how different weights in the network handle error and found distinct clusters for logic and memory.
Different facts, different fragility. Common facts like capitals stayed intact, while rarer details like company CEOs vanished almost completely.
Early step toward clean unlearning. The same technique might one day remove copyrighted or private data from models without breaking their reasoning.
The finding suggests that large models don’t “understand” math or memory the way we do. They replay facts from stored tables rather than compute them from first principles. In a way, these systems are students who remember the answers but never learned the lesson, and researchers are only now finding out where the cheat sheet is hidden.
TOOL OF THE DAY
AI models are finally learning to think in symbols. The new Nano-banana 2 model can now render full math solutions that look clean, coherent, and written like a human teacher on a whiteboard. This may sound small, but it’s one of the biggest leaps in how AI understands and expresses logic visually.
Here’s what’s happening:
From chaos to clarity: The first version of Nano-banana often mangled text and formulas. The new version writes math equations that actually make sense, line by line.
Context awareness: It knows it’s writing on a “whiteboard,” not just generating random text. The handwriting, formatting, and layout all look natural.
Smarter fusion: This happens because the language model (the “brain”) and the image model (the “eyes”) now work together. It is a precisely generated text, not a guess.
Why it matters: This means AI can now create accurate educational visuals like solved math problems, labeled diagrams, or clean infographic.
Try this yourself
Prompt an image model like Midjourney, Ideogram, or Playground to “write the full solution to a math problem on a whiteboard.” Notice how well it handles text. Compare it with earlier results and see how AI “understanding” evolves.
