Dan Luu vs. 'brain-off' coding, Meta's Muse caught routing to an OpenAI model, and Amp goes terminal-free

Amp's Mac app now runs agent threads with no terminal required. Amp, an AI coding agent, now starts a background 'runner' on your Mac automatically from its app, so long-running agent tasks keep going without you keeping a terminal window open. This is agentic tooling shedding its last bit of command-line friction — exactly the kind of change that lets non-engineer team members start using agents, which matters when you're advising a company on rollout. A 'Keep This Mac Awake' option stops the machine sleeping while the runner is on and plugged in (the screen still locks); on battery or with the lid closed it sleeps as usual, so it won't cook anyone's laptop in a bag. ampcode.com

Aikido open-sources a security model that runs on one server — built for air-gapped networks. Aikido Security released Altar-1, a free-to-download AI model specialised in security testing, slimmed down from Z.AI's GLM-5.3 so it can run entirely on-premises with no internet connection. Banks, hospitals and government clients who can't ship data to a cloud model can now run autonomous penetration testing in-house — and because the weights are public, security teams can inspect what the thing actually does rather than trust a vendor's claims. The slimming trick kept 168 of the model's 256 experts per layer with no retraining, and the result runs on a single node of four NVIDIA H200 GPUs — the kind of box a serious security team already owns. marktechpost.com

Meta ships Muse Spark 1.3 — built for agent work, with brakes on by default. Meta's new model release, aimed squarely at agent and coding tasks, with a 'max reasoning' mode for the hard ones. The interesting part is the behaviour, not the horsepower: it asks clarifying questions when your request is vague and checks with you before doing anything consequential — the cautious-by-default posture clients keep saying they want. In Meta's own engineers' comparisons it got jobs done with roughly 20% fewer tool calls and 25% fewer tokens than the previous version, and an open-weights release is on the roadmap. research.meta.ai

Meta's coding agent appears to quietly route some sessions to an OpenAI model. A developer investigating an odd session in his virtual machine found one of Meta Muse's subagent calls using a model labelled 'azure/muse-special', with encrypted payloads and call IDs in OpenAI's style. This is his inference, not confirmed by Meta: his best guess is an OpenAI model served through Azure. This turns 'which model is my agent actually using?' from a nerd question into a procurement question — and it comes with a privacy angle: unless you opt out, your Muse conversations can be used to train Meta's models. The agent's internal catalogue reportedly lists around 15 versions of Meta's own model alongside Claude, GPT-5.5/5.6 and Kimi K3 variants, so mixed routing looks like a designed feature rather than a one-off. mouse.dev

The Copilot email attack now has a CVE — and a setting you should check. You've already seen the Word-document worm from this disclosure series; the second instalment shows that hidden instructions inside an ordinary external email can steer Microsoft 365 Copilot in Outlook — enough to make it believe your inbox contains attack warnings, or to pull internal data into reply drafts. If a client runs Copilot against their mailbox, this is the practical follow-up to last week's story: the attack needs nothing but the target's email address, with the payload hidden as white text on a white background. Microsoft assigned it CVE-2026-55145 and mitigated it by sending external mail to a separate, less privileged agent before it reaches the main Copilot session — but customers have to enable the external sender flag in Exchange Online for that to work. enklypesalt.com

The NSA is spending billions a year just testing AI models. According to two sources familiar with classified estimates, the NSA told lawmakers it is spending billions of taxpayer dollars this year evaluating and testing frontier AI models, through its Artificial Intelligence Security Center, set up after high-profile hacking incidents. That's the sharpest public signal yet of how much 'knowing whether you can trust the model' actually costs — and it's reshaping the policy debate, with lawmakers concluding a comprehensive federal AI oversight system would run to tens of billions per year. The figure is significantly higher than previously known, and the administration has so far mostly resisted calls for greater federal oversight. washingtonsun.com