Symfony AI Mate and Claude: keeping AI on track with verified context
Dev news roundup 15/11/2026: AI agents, leaked secrets and updates

Dev news roundup 15/11/2026: AI agents, leaked secrets and updates

Several articles this week are about AI coding agents, and most of the news is bad: secrets published on GitHub, agents that a web page can manipulate, and deployment reports that turned out to be false. Jobs, technical debt and the quality of the web also come up. The week ends with several PHP releases to apply. Each source gets a reliability rating from 1 to 5.

AI coding agents: your secrets can leak without warning

Apiiro looked at Claude Code’s “Yes, and don’t ask again” option. When you accept it, the approved shell command is saved to .claude/settings.local.json, along with any passwords or tokens it contains. If that file gets committed, those secrets become public. According to Apiiro, 1,122 public files of this kind contained at least one secret. Press coverage gives slightly different numbers, so the original post is the one to trust.

Action: review which commands the agent can run without asking, keep the file out of Git, then scan your repositories. Rotate any secret you find: removing it from the repository is not enough, because it stays in the history.

echo ".claude/settings.local.json" >> .gitignore
git rm --cached .claude/settings.local.json
gitleaks detect --source .
trufflehog git file://. --only-verified

Also turn on GitHub secret scanning and push protection under Settings → Code security.

Reliability: 4/5. Apiiro specializes in application security, but it also sells a product that addresses this problem.

The WebMCP lab shows that an online store can offer an AI agent tools or instructions that push it to do things the user never asked for. Action: give browser agents as few permissions as possible and require human approval for any purchase or payment. Reliability: 3/5. It is a personal experiment, reproducible but not peer-reviewed.

A Medium post announces “948 Linux kernel vulnerabilities found by AI”. The tone is sensationalist and the article gives no verifiable details. Wait for the kernel maintainers to confirm before passing it on. Reliability: 2/5.

Trust your agent, but verify

A developer on dev.to describes how his agent reported “deployed, no 5xx errors”. In fact the containers were a day old: the commit had been pushed, but the server was never updated. Action: check the real state of the server instead of relying on the agent’s report.

docker ps --format '{{.Image}} {{.CreatedAt}}'
kubectl get pods -o wide
git rev-parse HEAD   # compare with the hash exposed by the app (/version, image label…)

Reliability: 3/5.

On cost, Hightouch describes a lightweight LLM “judge” that only calls Opus when it is unsure. According to the vendor, this gets close to Opus-level accuracy at a quarter of the cost. The pattern can help you supervise agents without a runaway bill. Reliability: 3/5. It is a vendor blog, and the result covers one specific use case, identity resolution.

AI, the developer job and the quality of the web

Jobs: in a post, Sylwia Lask points out that agents have not let anyone work less, despite the promises, and asks where that leaves developers. It is an opinion piece, not a study. Reliability: 2/5.

Technical debt: according to IEEE Spectrum, AI is finding flaws at scale, such as PHP2Shell, and that is forcing us to fix forty years of old code. Action: take stock of your legacy components and the PHP versions running in production (php -v, composer show --platform). Reliability: 5/5. IEEE Spectrum is the IEEE’s flagship magazine.

Effects on the web: on Lobsters, developers note that search engines rank more and more low-quality generated content. Two related reads: Camille Roux on what still counts as a work of art when AI writes the code, and a developer’s account of leaving Claude for Pi. Reliability: 3/5. These are discussions and personal opinions.

New releases: what to update first

Reliability: 5/5 (Packagist, GitHub and the official site).

Priority: apply the patch releases first (Symfony, Doctrine, FrankenPHP), since they usually don’t break anything. For the Laravel and Inertia minor releases, read the changelog before you upgrade.

composer outdated
composer update "symfony/*" doctrine/orm --with-dependencies
php bin/phpunit   # or: php artisan test

Other releases came up this week (Docker, Kubernetes, Node.js, Elasticsearch, Libghostty 1.0, Fedora 45). They are not in our sources, so we list them without version numbers or details.

Cyberattack at Crédit Mutuel Arkéa

According to the press, including Moneyvox and FrenchWeb, the bank confirmed in early November a cyberattack that enabled fraudulent transactions. The attackers reportedly got in through a service provider’s platform and used the bank’s own account. The funds were then reportedly converted into stablecoins, followed by a race to freeze them. The amounts reported vary from source to source, and the bank has confirmed none of them. Reliability: low for now. Wait for official statements.

Lesson for developers: detecting and blocking a transfer quickly matters as much as preventing it. Set up real-time logs and alerts on sensitive operations, and keep an eye on the access your service providers have.

What to watch

  • How the WebMCP standard proposal evolves, and what safeguards browsers add for agents (WebMCP lab).
  • Any official confirmation of the AI-found Linux kernel flaws, and the fixes for PHP2Shell.
  • Default settings in coding agents (confirmations, access to sensitive files such as .env) following the leaks reported by Apiiro.

Sources

Add a comment

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.