The week in AI: Funding caution, autonomous exploits, and JEPA theory

Artificial intelligence development faced a reality check this week as private investors grew more cautious toward massive infrastructure spending. At the same time, security researchers revealed new capabilities in open-weight models, and theoreticians published proofs regarding how different models learn internal representations of the world.## Industry and funding

OpenAI secured approximately $11 billion in a private funding round, falling short of its initial $30 billion target, according to reporting cited by The Decoder. The company originally targeted a $30 billion target at a $1.4 trillion valuation before lowering its goal to a $20 billion target at a $1.25 trillion valuation, as reported by TechCrunch. The shortfall highlights a broader shift in investor sentiment as Wall Street grows wary of heavy infrastructure spending after major technology firms reported weaker-than-expected quarterly results, according to TechCrunch. Major tech companies including Alphabet, Amazon, and Meta have committed hundreds of billions of dollars to data centers and network equipment, leading to negative free cash flow for several firms, notes TechCrunch. These market pressures contributed to a sharp Nasdaq index decline at the end of October, according to the BBC.

Models and safety

Researchers at Epoch AI reported that the open-weight AI model GLM 5.3, developed by Z.ai, can autonomously exploit Mozilla Firefox by targeting the patch gap—the window between a security fix becoming public and reaching users (unite.ai). Operating offline with zero human input, the model constructed multi-stage exploit chains across six Firefox versions in under 48 hours (the-decoder.com, unite.ai). In successful compromises, the model chained patch-gap vulnerabilities with unpublished sandbox-escape flaws to read local files, such as saved passwords and authentication cookies (the-decoder.com, unite.ai). The researchers noted that safety guardrails were not triggered by API providers during the exploit development (the-decoder.com, unite.ai).

Research

In theoretical machine learning, authors turned the Platonic Representation Hypothesis—the idea that various models converge to a shared statistical model of reality as they scale—into a formal theorem using Joint-Embedding Predictive Architectures (JEPAs) (arxiv.org, arxiv.org). The researchers proved that separate JEPAs trained on different modalities observing the same underlying world learn the exact same representation up to a single rotation, provided their positive pairs correspond to two distinct observations of the same scene (arxiv.org). Validating their theory across multiple datasets, the authors demonstrated that these good-view JEPAs were significantly more sample-efficient than standard next-token prediction used in language models (arxiv.org, arxiv.org).

What to watch

• Whether OpenAI will successfully reach its revised $20 billion funding target through retail offerings and family offices.
• How software developers and browser maintainers will mitigate risks arising from automated exploitation of the patch gap by open-weight models.