A Frightful & Buggy News Round-Up 📰
👋 Hello!
The clocks have gone back in Europe, Halloween falls on Saturday, and plenty of IT teams got their fright a week early.
In this edition, we take a look at:
🌐 Why two AI-found PHP bugs pushed governments across Europe to switch off their own websites
🔑 How a key deleted from GitHub, but never revoked, led to the LiteLLM supply chain attack
🏦 Why the UK's push to report AI "near-misses" could turn test logs into regulatory evidence
QUALITY FORGE 2026
A Big Thank You!
Just yesterday, we held our annual Quality Forge conference! 🎉
We had a great time discussing how AI is changing the playing field, together with so many amazing speakers from around the world. A huge thank you to our speakers, hosts, and of course - the audience! The viewers really pulled out all the stops with how active they were in the chat, and we got to feel the magic of hosting Quality Forge yet again. 🤩
If you didn’t get to make it to this year’s conference - don't worry! There’s a recording of it available for anyone on our page below. 🎥
TECH NEWS
🌐 AI-Found PHP Bugs Push Government Websites Offline Across Europe
Two critical flaws in PHP, the server-side language that Wiz says runs on roughly seven in ten websites, led governments across Europe to switch off their own sites last week. Known as PHP2Shell (CVE-2026-109618 and CVE-2026-109619, both rated 9.2 out of 10), the bugs sit in how PHP manages memory and let an attacker run code on a server without logging in. Security firm Calif found one with OpenAI's GPT-5.6 Sol model, and Google's Wiz found more attack paths using its Wiz Atlas agent alongside Google and OpenAI models. The Netherlands pulled its DigiD identity service and the lower house of parliament's website, and UK tech minister Baroness Lloyd told the House of Lords that 32 government services were unavailable, mostly as a precaution. Patches shipped on 21 October, yet Amazon says its honeypots caught an exploit attempt on 17 October, four days before the fix.
So why did so many organisations pull the plug instead of patching? Often they couldn't say where PHP was running. It hides inside control panels, shared hosting and products that never mention it, and Huntress researcher John Hammond warned that "the number of places you have to go looking to identify vulnerable PHP versions gets very big, very fast." A patched site can still be exposed through a third-party service that isn't, and Cloudflare watched exploit attempts climb from thousands before the announcement to millions after it. The organisations that stayed online could answer two questions within hours: where does this component run, and can we ship the fix without breaking anything? Regular security testing that keeps an inventory of runtimes and third-party services up to date, combined with a rehearsed emergency-patch process, makes that a day's work instead of a weekend offline.
🔑 A Deleted Commit and an Unrevoked Key Opened the Door to the LiteLLM Supply Chain Attack
LiteLLM is an open source gateway that lets apps talk to OpenAI, Anthropic, Gemini and other model providers through one interface, and it gets about 23 million downloads a week. On 21 October someone published a poisoned version, 1.107.0, to PyPI and Docker registries. It carried an infostealer that hunted for cloud credentials, SSH keys and AI provider API keys, plus a remote access trojan that started every time Python ran. PyPI quarantined it in under two hours, by which point Wiz counted about 49,000 downloads.
LiteLLM's postmortem traces the attack to a master key accidentally committed to a public GitHub repository in May. Developers removed the commit quickly but never revoked the key. On 10 October the attacker used it to reach LiteLLM's agent platform and extract a GitHub CLI token with owner-level access to every repository, then stole tokens from two maintainers, so later malicious changes looked like routine work by trusted engineers. LiteLLM calls the incident a "strong illustration of the risks associated with failing to rotate a potentially exposed key." In the same week, developer platform Railway told all users to rotate their keys after its own breach, which it believes was part of a wider campaign against AI inference providers.
Removing a secret from git history does nothing about the copies someone may already have made, so pipelines should treat any exposed key as stolen. In practice that means scanning commits for secrets before they leave a developer's machine, rotating automatically when one slips through, and checking that your CI/CD pipeline pins dependency versions, so a poisoned release can't flow into builds just because it's the newest.
FUTURE VISION
In Future Vision, we pick a story covering upcoming tech developments that you should keep an eye on:
This edition looks at how a UK regulator asking firms to report AI “near-misses” sounds like one more compliance form, but it could make test logs some of the most important records a company keeps.
Speaking at a Bloomberg event on 19 October, Jessica Rusu, chief data, information and intelligence officer at the Financial Conduct Authority, said the duty to report major incidents “must extend to significant AI failures and near-misses,” an idea borrowed from aviation, where, as researcher Steven Adler told MPs, near-miss reporting is mandatory in most jurisdictions. Only days later, Axios reported, in coverage summarised by TheStreet, that five major US banks, including JPMorgan Chase and Bank of America, were finding unauthorised AI agents in customer accounts, most of them logged in with the customer’s own password and some taking steps nobody had asked for. Aviation’s system works because everyone agrees what a near miss looks like, and AI has no such definition yet: is an agent that tried to move money and got blocked a near miss? If the FCA follows through, firms that already measure these events will help draw those lines, while firms that can’t detect a near miss will report none, which looks reassuring right up until the first real incident.
For QA teams, that changes what an agent test result means. A run in which an agent reaches beyond its scope and gets stopped might still pass, yet it is exactly the event a regulator may soon want counted. A sensible first step is to define what a near miss is for your product, then make test harnesses and production monitoring record those events in the same format. Good AI testing should also check that your systems can tell an agent from a human and keep an agent to read-only access when that is all it was given, so the numbers exist before anyone asks for them.
HOT OFF THE PRESS
Get in-depth insights by reading up on the latest from our engineering experts
The Importance of Integrating Security Testing into Your CI/CD Pipeline
Fintech Testing: Ensuring Security, Reliability, and Accessibility in Financial Platforms
Want to see more? Check out our blog by clicking the button below.
CAREERS
We have 6 open vacancies for seasoned specialists across different teams. We are also hiring engineers outside Latvia, so please don’t hesitate to apply.
You can find more information about our open positions on our careers page, where you can also apply. If you think you have something to offer and want to be a part of something amazing — we would love to see your CV!
BUGS, BUGS, BUGS EVERYWHERE
TestDevLab SIA · Vingrotaju 1 · Riga LV-1010 · Latvia