Tomorrow, in context.

A 4-bit GPT-OSS model has, by Multiverse Computing's own numbers, beaten its 16-bit source on 7 of 9 benchmarks -- a result that inverts the basic assumption every deployment engineer works from, self-reported and unverified outside the company's own tests. NVIDIA is retiring the static benchmarks the chip industry has used for years, replacing them with session-replay testing that shows coding agents burning roughly 15 times more tokens than chat -- then grading its own next-generation chips against that new number. Meanwhile, a small MCP agent listing from Botskillstack, built only to look up Brazilian census data, ships with dry-run previews before every write and a firewall against instructions hidden in the data it reads. Three infrastructure claims define this week, and all three ask you to trust the entity making the claim. None of these vendors are lying. But none of them are neutral either, and the gap between the two is where this week's real story sits.

• •Multiverse Computing says its 4-bit compression of GPT-OSS outperforms the 16-bit original, a claim strong enough that it needs testing beyond the company's own benchmarks.
• •A Botskillstack MCP agent listing for looking up Brazilian census data ships with mandatory dry-run previews and content quarantine, a pattern worth copying for any agent with write access to real systems.
• •NVIDIA is retiring static prompt benchmarks in favor of session-replay testing that shows coding agents burning roughly 15x more tokens than chat, then using that number to grade its own upcoming chips.
• •An independent researcher got an AI agent to produce a working cohttp exploit in under a minute, collapsing the lag time defenders used to count on after disclosure.
• •Anthropic launched a community plugin marketplace for Claude Code with mandatory automated security scans and nightly catalog refreshes, replacing ad hoc GitHub trust as the default distribution channel.