AI agents find exploits from bug rumours alone; OpenAI blocks Cursor after SpaceX acquisition; GLM-5.3's 753B weights go open.

Latent Space

[AINews] OpenAI Shuts Off Cursor

OpenAI cut off Cursor's access to its models after SpaceX acquired Cursor. OpenAI cited that Elon Musk's companies violated contracts. Cursor said OpenAI models were only 5 percent of its traffic. This action reflects tension between OpenAI and Musk's AI division. The incident shows both companies have sufficient AI power to make access restrictions matter.

Z.ai / Hugging Face

GLM-5.3 Is Now Open-Weight

Z.ai released GLM-5.3, an open-weights language model with 753 billion parameters. The model shows 50 percent improvement on coding tasks over GLM-5.2 and provides state-of-the-art cybersecurity capabilities for vulnerability discovery. GLM-5.3 supports 1 million token context length and multiple serving frameworks.

Chips and Cheese

Hot Chips 2026: Samsung's Processing-in-Memory (PIM)

Samsung embeds compute units into LPDDR5X memory banks. The design uses multiply-accumulate units to perform computation within memory. This provides 614 GB/s internal bandwidth, much higher than standard LPDDR5X. Software constraints limit practical use: the chip must change memory modes to compute, which stops multitasking, and uncacheable memory reduces performance. Eight chips together equal Intel Meteor Lake NPU throughput.

pwning.systems

I Accidentally Turned LLM Memory into Program Analysis

The author built Lemmalog, a Datalog-based system that manages LLM memory for vulnerability research. The system uses structured knowledge and logical rules instead of traditional retrieval. It tracks why conclusions exist and removes dependent facts when assumptions change. Lemmalog reduces context size 38 times on benchmarks compared to standard approaches. The system treats vulnerability investigation like program analysis, where LLMs handle natural language and the database handles logical reasoning.