LLMjacking wave hits AI dev platforms — rotate your API keys now
Wiz counts nine breached AI dev platforms in October, Northflank the latest — plus ChatGPT folds Work into chat, a VLM accuracy trick, and AI in the kernel.
Copy markdown
Northflank is the ninth: 'Black October' for AI dev platforms
Wiz has now logged at least nine breaches of AI developer platforms in October, dubbing it 'Black October'; Northflank (disclosed Oct 30) told users to assume every environment variable was decrypted, and customers confirmed their AI keys were abused. In parallel LLMjacking cases Wiz tracked, two stolen cloud keys ran ~800K and ~1.7M calls — mostly to Claude — for an estimated $100K+ on the victims' bills. Rotate AI and payment keys first, scope new keys by model and region, and set spend caps today.
ChatGPT folds 'Work' back into the main chat
OpenAI is merging ChatGPT's separate Work tab into the main desktop chat, rolling out to select users first; Codex stays its own app. It follows December's move to pool ChatGPT and Codex into one usage budget — so your projects and conversations live on one surface again.
Fix the image, not the prompt, for vision models
An EMNLP 2026 study on the SPaRC spatial-puzzle benchmark found annotating the image beats prompt tweaks by a mile: adding grid labels (like 'S' and 'E') lifted Qwen 3.5 397B from 14.4% to 32.6%, and up to 48.4% with the best image version, while text edits moved o4-mini at most 5.2 points. If your VLM app stumbles on spatial tasks, pre-label the pixels instead of rewording the prompt.
Linux 7.4-rc1 is the biggest merge window ever — and AI is in it
Torvalds cut 7.4-rc1 with 19,671 commits, the largest merge window in kernel history (21% over 7.3), and commits declaring AI assistance roughly 2.4x'd. He said he used Claude to prototype a file-read optimization, which surfaced three real bugs — a signal that AI-assisted patches are now mainstream in the most scrutinized codebase around.
YC Labs set agent swarms loose in Minecraft and Mindustry
Y Combinator's labs ran fleets of LLM agents in Minecraft and Mindustry to study coordination at scale — and reported the swarms struggle to self-organize, 'drowning in bureaucracy' rather than building a working society. A concrete, reproducible look at where multi-agent setups break before you wire one into production.