This week the risk in AI automation shifted from the models to the tooling around them: extensions, desktop apps and agents leaked credentials and data. Model prices kept falling, so most of the cost now sits in how you build the workflow.
A trusted extension turned malicious
Anthropic's verified Cursor account pushed a malicious 2.1.313 build of the Claude Code extension. It stayed live for nearly 12 hours, auto-installed on around 5,000 developer machines and hunted for cloud credentials. If your team or a contractor uses Cursor, check the time window in my post on the backdoored Claude extension on Cursor, rotate SSH keys and cloud tokens, and update to 2.1.314.
Postman rewires your coding agents
Postman 12.33.4 silently installs Codex and Claude Code plugins on launch and adds a hook that routes API work through Postman skills. Any machine where you build automations with agents needs a quick audit of ~/.codex/config.toml and ~/.claude/settings.json. The cleanup steps are in how Postman auto-installs AI plugins into your agents.
An agent that shared your email
An open GitHub issue labeled as a security bug says Claude Code put the user's account email into the User-Agent header, so every site it visited could read it. The reporter says they now get dozens of scam emails a day. My takeaway in the post on the Claude Code email leak: run agents on business accounts under a separate service identity and log what they send out.
Usage up, AI spend flat
Ramp data shows the median company spends $12-15 per employee per month on AI, and spend grew only 1.5% across August and September while token volumes kept climbing. Price cuts and routing simple tasks to cheaper models drive this. As I wrote in what Ramp's spending numbers mean for small business, most of the budget goes into building and connecting the workflow, not tokens.
Retrieval beats a bigger model
Box is building Nvidia Nemotron embeddings into a retrieval layer so agents read only the relevant chunks instead of whole documents. Nvidia claims 90% fewer tokens, but those are vendor numbers, and the feature reaches only Enterprise Advance customers early next year. The lesson in my breakdown of Box and Nemotron retrieval applies without Box: if your agent reads a whole contract to answer one question, fix search first.
An open model for document fields
Dharma-AI released Dharma Field Extractor, a 2B-parameter open model that takes a document image plus a schema and returns typed JSON. For tender, invoice and contract pipelines, that means parties, deadlines and penalties come back in the shape your system expects in a single call. I'd test it on a sample of real documents first; details in the Dharma Field Extractor post.
What to do this week: check whether anyone on your team ran the Claude extension on Cursor during the affected window, and rotate keys if so. Audit agent config files on every machine that runs Postman. Move any agent that touches business data off personal accounts and onto a separate service identity.