AI engineering, October 31, 2026

Top News

GPT-6 Rosalind is identified as gpt-6-rosalind. If you want to use it for internal research, the starting point is the trusted-access program.

Here is how access works:

• Organization approval is part of eligibility, so access is tied to the organization conducting the research.
• User authorization is required as well. Before using the model for internal research, check your organization’s approval and your own authorization. The access rules require an approved organization and an authorized user for internal research.
• Internal research is the permitted scope of use, rather than unrestricted authorization for other work.

One catch: use is limited to approved organizations and authorized users conducting internal research.

Try it: Request access to gpt-6-rosalind through the trusted-access program for your organization’s internal research.

Sources: developers.openai.com

Top News · 855 HN points

If you installed or updated the affected Claude Code extension in Cursor during the October 29 incident window, check your machine for compromise. The maliciously modified package collected host and user information and attempted to install a persistent SSH backdoor.

Here is what affected users should do:

• Check whether your macOS or Windows installation falls within the affected versions and time window.
• Investigate telemetry, isolate affected hosts, and remove persistence using the platform-specific remediation scripts.
• Rotate credentials from a trusted system, not an affected machine.
• Distinguish the modified package from Anthropic’s work. The malicious versions were not developed by Anthropic.

One catch: this notice covers only the identified package. Cursor reported no evidence that other versions or the official distribution were affected.

Cursor said it removed the malicious versions and rebuilt the marketplace. It also said the attacker server was taken offline, disabling the reverse shell and host-detail reporting.

Sources: cursor.com

Top News

Baseten’s previous RL rollout update path paused serving and kept a full checkpoint copy in host memory. Baseten’s delta weight-sync system applies sparse updates directly to serving GPUs, dropping that host-memory copy.

Here is what changed:

• GLM-5.3 sync time fell from 36 to 15 seconds, while the request pause dropped from 6 to 0.8 seconds.
• Removing the CPU checkpoint copy avoids roughly 770 GB of host-memory use on GLM-5.3 and lets the next delta apply when it arrives.
• Checksum verification rejects corrupted downloads or a wrong parent before GPU writes. A Kimi K3 30 GB delta improved from about 230 to 100 seconds.
• Host-memory KV offloading lifted AgentX cache hits from 2% to 92% at concurrency 48, with output throughput per GPU rising 3.2×.

One catch: sync timings used one replica and trainer-shaped fixture deltas. The AgentX throughput gain disappeared at concurrency 16 because the KV cache fit on the GPU.

Sources: baseten.co

1. Amazon SageMaker AI adds up to seven isolated partitions per GPU for inference
2. Baseten adds joey-1, an open-source 2B decision model for classification, routing, and scoring
3. Kioxia memory sales grow on data center demand as chip prices rise · Reported · nikkei.com
4. “A Fix for Softmax and Cross-Entropy” eliminates two numerical bugs in training gradients
5. Superpowers framework adds test-driven development and subagent code review to coding agents · 33 Stars today
6. Microsoft’s 4B Mage-VL-Cowork computer-use model hits 66.0% on OSWorld-Verified with a 50-step budget · 37 Likes