An autonomously self-propagating AI virus has been demonstrated
A joint team from the University of Toronto, the Vector Institute, the University of Cambridge, and ServiceNow has released a proof of concept (PoC) for a "computer virus that keeps spreading autonomously," built by abusing open-weight LLMs. The mechanism runs an LLM on GPUs it has seized and uses inference to adapt its attack methods to each target as it spreads; because it relies on no API, the researchers say, it is also hard to monitor and neutralize.
In their experiments, the virus achieved roughly 80% success at detecting vulnerabilities, about 53% at successfully exploiting them, and 88% at self-replication, for an overall success rate of around 37%. The team states that "self-sustaining AI-driven threats are no longer merely theoretical," and points to the possibility that the future internet could become an "ecosystem" where offensive and defensive AI agents intermingle.
AI is a "good engineer" but a "bad researcher"
Also drawing attention were the results of an experiment asking whether AI can do real research. Teams from Princeton University, the AI Safety Institute, and elsewhere had frontier agents tackle the "core questions" of two unpublished papers submitted to NeurIPS 2026, then ran a "shadow evaluation" in which the papers' own authors reviewed the agents' output.
The result: the agents could handle the engineering work involved, but could not produce research original enough for a top conference. They reportedly locked into a small number of directions early on, responded inadequately to feedback, and tended to stay stuck on unpromising lines of attack. For both papers, the authors rated their output as a Reject. This looks like an important clue for gauging how quickly "recursive self-improvement" — AI autonomously driving its own R&D — might progress.
More than 1,337 AI industry figures ask the US for "pacing" of AI development
A statement has been released, signed by chief scientists and CEOs from major labs including OpenAI, Anthropic, Google DeepMind, Thinking Machines, Meta, and Safe Superintelligence Inc. Citing the possibility that "we may be approaching the automation of AI research," it asks the US government to help build an international framework for developing the technical and governance tools needed to deliberately modulate progress across the frontier.
Given that no single company or country can now afford to slow down unilaterally, the statement argues for mechanisms that buy society time to adapt at each new threshold of capability. Notably, its focus is less on safety per se than on the problem of how to design human coordination and collective action.
NVIDIA's CUDA is being threatened by AI coding agents
Reports say the CUDA ecosystem — the source of the competitive advantage NVIDIA has built in the GPU market — is facing new pressure from AI coding agents. If agents that automatically rewrite and port CUDA-equivalent code for other vendors' GPUs spread widely, NVIDIA's CUDA lock-in could lose its grip.
With the hardware performance gap narrowing, the software-layer moat has been regarded as the pillar of NVIDIA's advantage — which makes the question of how coding agents erode that moat a perspective that could shape the outcome of the GPU race.
Quantization damages knowledge "nonlinearly"
A case study using Qwen3.6 27B has published an analysis concluding that quantization — the standard technique for running LLMs locally — damages knowledge nonlinearly: dropping the bit depth by a single level has limited impact, but past a certain threshold, capability falls off sharply.
Meanwhile, hands-on reports on Reddit note that DeepSeek V4-Flash-0731 "gets hit by quantization like a truck": at Q3 it can serve as a Qwen3.6-27B-class alternative, but at Q2 its practical usefulness drops. That quantization tolerance varies this much from model to model has once again been thrown into sharp relief as a reality of running models locally.
The EU opens bidding for seven AI "gigafactory" sites
The European Union has started accepting bids for seven "AI gigafactories," a roughly €30 billion initiative aimed at restoring Europe's AI competitiveness. The goal is to build large-scale compute infrastructure for training inside the region and correct its lag behind the US and China.
Challenges remain around reconciling the plan with decarbonization and securing electricity, but the EU appears to be starting to steer away from a strategy of butting heads over regulation (the AI Act) toward carving out its own path on compute as well.