18 Nov 2026
Caledapt Signal | 18 Nov 2026 - Agent safety advances amid rising funding and model releases
Scottish SMEs exploring AI agents should prioritise isolation techniques that limit potential harm from errors or misuse. Recent work highlights practical paths to safer experimentation without requiring enterprise infrastructure.
Today's AI Signal
External sandboxes for coding agents explored
What happened. Simon Willison reports experimenting with external sandboxes to run code from agents, aiming to make it safer to test local models by reducing the risk that mistakes or prompt injection lead to damaging actions on a machine.
What Caledapt recommends: Caledapt recommends testing agent code execution in isolated environments such as containers or restricted user accounts before broader use.
Read the External sandboxes for coding report · Published 17 Nov 2026
Datasette Lite updated to run full server-side app in browser
What happened. Simon Willison released a rewrite of Datasette Lite using service workers and Pyodide, allowing the full Datasette Python application to run in the browser via WebAssembly, closely matching the server-side experience.
What Caledapt recommends: Caledapt advises using Datasette Lite for internal data exploration with non-sensitive data, avoiding transmission to external servers during evaluation.
Read the Running the full server-side report · Published 17 Nov 2026
The Future Build Lab
Beginner
Agent Log Policy Checker
The problem: Agent logs may contain records of actions that violate operational policies, such as attempts to write files outside approved directories, which can go unnoticed during manual review.
The solution: A script that reads a structured agent log file and checks each entry against a defined policy, outputting only policy violations without executing any agent code.
Example input: 2026-11-18T10:32:0Z agent_action write_file /home/user/project/data/../tmp/test.txt
Expected output: Policy violation flagged: the log records an attempted write outside the approved workspace at 10:32.
Stack: Python 3.9+, standard library only (json, re, datetime)
Why it matters: This lab provides Scottish SMEs with a auditable way to review agent behaviour post-hoc, supporting oversight without introducing runtime risks from untrusted code execution.
First step: Create a sample log file in JSON lines format containing one permitted and one prohibited file write action, then write a Python script to read it and flag entries where the target path falls outside a permitted prefix.
Tools Worth Exploring
langchain-core
Version 1.8.2 adds headers to langsmith_gateway and proxy ignore functionality.
Why try it: The update improves configurability for LangChain-based agent workflows used by Scottish SMEs integrating with LangSmith while operating behind restrictive network policies.
View the langchain-core 1.8.2 release notes · 1.8.2 · Released 17 Nov 2026
One Thing to Try Today
Want to make AI useful in your business?
Book a free 30-minute consultation with Caledapt.