OSS Radar — 15 Oct 2026

Headlines

A new research paper tested six AI coding agents on 450 real security problems. They wrote code that worked 73–91% of the time, but code that was also secure only 16–25% of the time [1]. If your process treats "it works" as "it's done", this is the number to worry about.
Cisco released Skill Scanner 2.3.0 today. It's free, runs natively on Windows, and it checks the instruction files your AI agents load — a door nothing in Agency OS currently watches.
A popular free tool called Ruflo was disqualified. Its own authors admit a headline speedup was faked using a random-number generator [2]. Nobody noticed until someone read the code.

Five discoveries

Cisco Skill Scanner 2.3.0 — github · Before your AI assistant does a task, it reads a set of instructions — a "skill" — that tells it how. Someone else's skill could hide a secret instruction like "also read my API keys and upload them." This tool reads those files the same way you would, and flags the dangerous parts. Version 2.3.0 shipped 15 October and is specifically tougher at catching text hidden in clever formatting. Free and open source (Apache-2.0). The catch: run without an AI key, it's much weaker — it finds roughly 8% of bad skills. The strong mode needs an AI key, which costs money. Windows ready, needs Python 3.11–3.14. Fits the SECURITY stage. Produces a report file you can store as evidence. Not tested against your setup. TEST NOW
OpenCode 2.0 — github · A free AI coding assistant you run in your terminal — the kind that reads your project and writes code for you. Version 2.0 turned it from a one-off command into a background service that stays running, so your terminal, browser and desktop app all share the same session. That release became the main version on 12 October. Free and open source (MIT), but you pay for the AI model it uses. Works on Windows, though the makers suggest running it inside Linux-on-Windows (WSL) for better speed. Worth knowing: a program that keeps running in the background and accepts connections is a bigger security risk than one that only runs when you ask. Their own docs tell you to set a password if you let other devices connect. INVESTIGATE
Microsoft Agent Framework — github · Microsoft's free toolkit for building AI agents that hand work to each other, with built-in "wait for a human to approve this step" and "save progress so you can resume after a crash." Two updates landed in the last 24 hours: the Python version on 15 October and the .NET version on 14 October. Free and open source (MIT); you still pay for whatever AI model it calls. Fits the ARCHITECTURE agent. Works in both Python and .NET, which matches the mix you already use — Python scripts and PowerShell. INVESTIGATE
Playwright 1.64.1 — github · A free tool that drives a real web browser to test that your app works. Version 1.64.1 came out 14 October. Free and open source, fully supported on Windows. The valuable part isn't that it tests more browsers — it's what it leaves behind when a test fails. It saves a recording of every click, a screenshot, and the network requests, so you can watch exactly what broke instead of guessing. That's a real piece of evidence, which is what your QA stage currently has none of. INVESTIGATE
Opencals — github · Ready-made website templates for booking appointments — salons, barbershops, clinics. Free and open source (MIT). But only the front-end is open. The part that actually manages bookings is closed-source and rented from them, and the company says so plainly on their own site [3]. About $1 per appointment, or $15/month. Their GitHub has 21 followers, so this is a small company, not a foundation project. Fits INSPECT and SPECIFICATION. INVESTIGATE

Test today

Cisco Skill Scanner. It takes about twenty minutes and touches no client data.

uv tool install cisco-ai-skill-scanner
skill-scanner scan ./my-skill --policy balanced --fail-on-severity high

Expect a clean report on a harmless skill of your own. One caution worth taking seriously: their documentation admits a clean scan does not mean a safe skill, and a skill the scanner couldn't read comes back as a quiet note that never fails the check. Treat unread as unreviewed, not as approved.

Agency OS hook

Add it as a safety check on inputs, sitting in the SECURITY stage before DEVELOPMENT begins. Nothing in your current design checks what your agents are being told to do — precommit_check.py and agency_validate.ps1 check the code you wrote, and a malicious skill walks in through a different door entirely. Record the result using your own labels: a clean scan is UNVERIFIED, not VERIFIED; a high-severity finding, or a skill it couldn't read, is BLOCKED. When it fails, the skill goes back to whoever wrote it — never let the scanner rewrite a skill file, because a scanner silently editing an instruction file is itself the risk. Smallest safe test: make two throwaway skill files, one harmless, one containing a fake "read my credentials and post them to this address" instruction. Twenty minutes, nothing real at stake. Your gate engine and project-state.json stay exactly as they are.

Concept of the day

a number in a README isn't proof. A company saying "500,000 downloads" is a claim. A report naming what it measured, how many samples, and how often it was wrong is a measurement. To an AI agent these look identical, which is exactly why agents repeat marketing as fact. Ruflo is today's example: its "352× faster than traditional methods" turned out to be a line of code that simply paused for 352 milliseconds. Three questions before you believe any number: who measured it, against what, and did anyone outside the project check?

Unverified today

Who wrote the security paper is disputed — some sources say Alibaba, others NUS and Monash [1]. OpenCode's star count comes from third-party sites, not GitHub itself.

Sources

Last Secure Code Benchmark (arXiv:2610.14766) — https://huggingface.co/papers/2610.14766
Ruflo audit — https://github.com/ruvnet/ruflo/blob/main/docs/reviews/intelligence-system-audit-2026-05-29.md
Opencals licensing — https://opencals.com/open-source