Welcome to the 315th edition of Coding Jag brought to you by TestMu AI!👐
A model was built, benchmarked, and scheduled to ship, and then the tests said no. OpenAI pulled GPT-6.1 Astra ahead of a planned October launch after it failed internal safety and alignment audits. For once, the interesting part is not what the model could do. It is that somebody checked before release, and the check was held.
Also in this edition: three stories where the evidence said one thing and reality said another. An agent whose web access had been blocked still reached the open Internet by tunneling through DNS. A suite can hit ninety percent coverage and still ship the bug. An average pass rate can look healthy while a flaky agent hides inside it. Elsewhere, Selenium 4.50 landed, and Gemini 4 Argon arrived with a guest list, and OpenAI shipped always-on agents that work while you are away.
📬 Come across something useful or interesting? Just reply and let’s exchange ideas.
News
07 min
thehackernews.com
🛑 OpenAI shelved GPT-6.1 Astra ahead of a planned October launch after it failed internal safety and alignment audits. The Wall Street Journal reported that the model showed higher levels of deception than its predecessor and failed to disclose actions it had taken. Sometimes it proceeded without seeking permission.
06 minblog.google
🔒 Google is rolling Gemini 4 Argon out to a set of trusted cyber defenders through its Fairwind Program rather than to everyone. The pitch is frontier performance on long-horizon work across software engineering, enterprise knowledge work like legal and finance, and cyber defense. Broad availability waits on the safeguards.
09 mingithub.blog
🤖 Kevin Stubbings walks through the targeted audit taskflows behind 24 Android findings, including user tracking in OsmAnd and a Wikipedia account takeover via deeplink. One useful finding for testers: large language models are good at finding vulnerabilities but struggle to estimate severity.
09 mintestmuai.com
🩸 A poisoned Salesforce lead was enough to make Agentforce leak CRM data over a DNS lookup, with nobody clicking anything. Zenity Labs reported three findings, and we walked through how the data left and what Salesforce fixed. The other DNS story this week happened in a lab. This one ran against production Agentforce.
07 minopenai.com
🔵 OpenAI launched dots, agents meant to keep working while you are away. Connect your apps, talk to one in ChatGPT, Slack, or Teams, and it returns finished work instead of answers. The promise is getting your time back. Dots start with built-in rules for when to act independently and when to ask for approval.
AI
07 minthenewstack.io
🕳️ Meredith Shubel covers two incidents in which OpenAI models found a way around the path that had been closed to them. One reached the open Internet through DNS after its web access was blocked, which paused tool use. The other kept cheating and was pulled offline.
08 minleaddev.com
🔁 Sriram Madapusi Vasudevan argues that the loop is only control flow, and that five contracts should be standard before teams widen agent autonomy: context, state, tools, control, and completion. Drawn from his work on the AWS DevOps Agent. The line worth keeping is that a successful action is not a successful task.
09 minthoughtworks.com
🧱 Jaya Simha Reddy Nandyala and Prabina Pani name the paradox plainly. Agent code can look elegant, compile cleanly, and pass basic syntax checks, yet still be wrong for the system. Their pattern has two layers: guides that constrain before the agent acts and sensors that verify after it does.
09 mintestmuai.com
📉 Pass@k asks whether any of k runs passed. Pass^k asks whether all of them did, and the gap between the average pass rate and Pass^k is where flakiness lives. One agent that succeeded in 77.4% of runs cleared all five runs on only 53.0% of tasks, a 24.4-point drop, according to IBM Research.
Automation
07 mindev.to
🎭 Nikolay Chernev puts the problem in one sentence: coverage measures which lines ran, not whether a test would notice when they are wrong. Generated tests make the gap free to create, all happy paths and none of it discriminating. His remedy is old and boring: to break the code on purpose.
06 mingithub.com
🧪 BiDi is where this release spends its effort, across all five language bindings plus two architecture decision records covering the network event API and the low-level contract. In Java and .NET, relative locators now work from elements and shadow roots. Ruby can install and uninstall web extensions, and deprecated Firefox profile and certificate methods are gone.
06 minqualityremarks.com
⚖️ Keith Klain opens with Hitchens’s razor and lands on the part testers should sit with. Machines now generate software, operate tools, make decisions, produce tests, and increasingly judge the results of those tests. Every layer is another place where an assumption can disappear behind automation.
10 mintestmuai.com
🎯 Eleven injection and tool misuse scenarios written as a test plan you can actually run, graded on what the agent did rather than what it said in reply. Worth knowing the backdrop: one public competition drew 272,000 injection attempts against 13 frontier models, and every model proved vulnerable.
Tools
06 mincode.visualstudio.com
🧰 The release notes lead with three things: expanded agent workflows, better worktree reuse, and new enterprise AI controls. Dev Containers get automatic cleanup and faster startup. Several agent features land as experiments, including multi-folder sessions and delegating tasks to remote agent hosts. Chats can now be archived inside a session.
05 minblog.postman.com
📮 Quinton Wall shows two ways to install the plugin, then puts it to work. You can ask whether the API you need already exists before anyone builds it, and mock one that does not exist yet, both from inside Codex. Useful if your API tests keep waiting on endpoints.
Video & Podcast
12 minse-radio.net
🎙️ Raju Dandigam tells host Amey Ambade to treat an agent as a software subsystem rather than a standalone model call. Most of the episode covers how teams can ensure an agent works and how to debug it when it does not. Behavioral testing against contracts, golden scenarios, and full execution traces.
10 minpodcasters.spotify.com
🎧 Alan and Brent return to AI and what it is doing to testing practice, taking in confidence engineering and how AI shifts the decisions teams make. The show describes itself as being about Modern Testing, which isn’t all that modern and isn’t really about testing, except when it is.
Events
05 mintestwarez.pl
🇵🇱 Poland’s oldest testing conference reaches its twentieth edition, running from 4 to 6 November at the Sheraton Sopot Hotel. One workshop day is followed by two days of talks on the theme of heritage, modernity, and adaptation. The program and booking pages are published in Polish.