self.md radar — 2026-10-09
the evidence has a stop button
Screenpipe’s new MCP release says its protocol tests do not establish whether an answer is good. It also adds a source trail to retrieval. Microsoft has made a containment SDK stable; Anthropic has specified what a physical agent must do when its model connection drops.
1. Screenpipe MCP 0.20.3 returns less by default
Screenpipe’s October 8 MCP release
now returns source IDs and links with search results. Returned text is bounded by default; full text requires max_content_length: 0. Pagination no longer suggests another page after the result set ends.
That gives an agent a route back to the original recording and makes the larger retrieval request explicit. Helper processes now exit when their client disconnects, including when the backend request has stalled.
The release reports 156 package tests and 30 deterministic protocol checks. It also says those checks do not measure answer quality or production outcomes.
reading: Screenpipe MCP 0.20.3 release
2. MXC 1.0 stabilises a cross-platform sandbox API
Microsoft has shipped MXC SDK v1.0.0 , naming its Rust, .NET, and Node interfaces the stable compatibility boundary for the 1.x line. The SDK runs untrusted model output, plugins, and tools behind a shared container-policy model on Windows, Linux, and macOS.
Its Node example begins with network egress denied. That is only the start: MXC exposes filesystem, network, and UI rules, and its documentation expects an application team to tune them when a workload hits an access denial.
Audit mode switches sandbox security off for the workload under analysis. The docs label that plainly. So should any team that uses it.
reading: MXC 1.0 release · MXC repository and policy documentation
3. Anthropic’s new policy specifies a safe state for hardware
Anthropic’s updated Usage Policy , published October 8 and effective November 12, groups fake accounts, fabricated sites, hidden sponsors, and related influence machinery under a prohibition on deceptive campaigns or artificial activity.
Anthropic says it removed its blanket ban on personalised vote and campaign targeting while retaining restrictions on deceiving voters or disrupting elections. The update cites multilingual voter information and ballot-cure notices as legitimate civic uses.
For autonomous physical action, it requires a qualified operator who can observe and stop equipment, plus a safe state if Claude disconnects. That is a product requirement with a policy citation.
reading: Anthropic’s policy update · Usage Policy