Does your website load a script from another vendor, for A/B testing, analytics or a chat window? Does a service provider hold your customers’ names and addresses? Does an AI tool have access to your servers, or could it obtain that access? In November 2026, each of these became a way in for an attack.
On 14 November, the French postal service La Poste announced that the names and addresses of 22 million people had been stolen from Maileva, its subsidiary that sends mail on behalf of businesses. On 16 November, an attacker slipped their own code into Optimizely’s A/B testing script, loaded by the websites of at least 178 organisations. The same day, Google described an attack tool that compares a piece of software’s code before and after a fix, then turns the difference into an attack, with no human involved.
We draw one rule from this for businesses that have their software developed: nothing reaches production without going through the deployment pipeline, neither another vendor’s script nor an AI.
The technical details are grouped in the “For your technical team” box.
Recent cyber attacks: what happened in November 2026
In October, attackers had taken remote control of refrigeration controllers in supermarkets in France. Overnight on 1–2 November, new attacks hit shops in Belgium, the Netherlands, Sweden, Luxembourg, Ireland and Estonia.
On 2 November, in France, after an incident at a supplier of refrigeration equipment, temperature thresholds were changed in the monitoring of the cold rooms at three sites of the Établissement français du sang, the French national blood service; the alarms went off and no product was lost.
On 4 November, the French bank Crédit Mutuel Arkéa announced a cyber attack. According to the bank, the attackers went through the platform of one of its service providers and used the bank’s own account to make transfers to other banks; its customers’ funds were not affected. The next day, the bank put the amount diverted at more than €100 million.
On 12 and 13 November, Outlook, Teams, SharePoint and Copilot were unavailable to many Microsoft 365 users around the world. Nothing shows that it was an attack: Microsoft describes “a technical issue not related to cyber or AI / agents”, which appeared after a change to its network.
On 13 November, the FBI published an alert on ARTEX, an open source tool built for testing your own security, which lets AI agents carry out the stages of an attack on their own, and which attackers are using (details below). Ten days earlier, in its Annual Review 2026, the NCSC had set out a pessimistic scenario for 2028, “The age of autonomous attack”, in which advanced AI attack capability reaches “inexperienced cyber actors”, a scenario whose purpose, it says, is “to illustrate the consequences of inaction”.
A hijacked third-party script: protecting your site without breaking A/B testing
Optimizely sells an A/B testing tool: on every page, the site loads a script from the vendor that shows each visitor one of the variants being tested. The script is downloaded from Optimizely’s servers on every visit. It also contains the code the customer writes for their tests, in Optimizely’s interface.
On 16 November, an attacker exploited a flaw in Optimizely’s sign-up process to impersonate users and write their own code into this script, for some of its customers. Optimizely closed the flaw and removed the code that same evening. In the meantime, the code ran in visitors’ browsers. According to Report URI, a website security company, it replaced cryptocurrency wallet addresses with the attacker’s and copied the card numbers typed into the sites’ forms. Report URI found it in 188 of the 1,611 scripts it examined, at 178 organisations, including the BBC, CNN, Zoom and HSBC (SecurityWeek).
None of these organisations had changed their site. The attacker’s code reached production without going through their deployment pipeline, their tests or a review.
Two common protections fall short here. A Content Security Policy (CSP), a header that tells the browser which scripts it may run, protected the sites that had a strict one: the injected code loaded the rest of the attack from another domain, which their policy did not allow. According to Report URI, allowing Optimizely’s domain was not enough: “What separated the 24 protected sites from the rest was a restrictive script-src”, the part of the CSP that lists the allowed scripts. A CSP therefore limits the damage. It does not stop code slipped into a script from a domain it allows.
Subresource Integrity (SRI) makes the browser refuse a file that no longer matches the hash written into the page. It does not suit this script: the part that holds the tests “changes frequently as you create, update, and conclude experiments”, Optimizely writes, and cannot be checked by simply comparing a hash. A fixed hash would block the script at the first test that changed.
Two approaches keep A/B testing without bypassing the pipeline.
Choose the variant on the server. With Feature Experimentation, Optimizely’s product for developers, the decision to show one version or another is made by the application itself: the software development kit is installed on the server, and “feature decisions are made on the server”. This kit is a dependency like any other: it is installed at its exact version and goes through the pipeline. Each version being tested is then application code, switched on by a configuration setting: every test goes through the developers, and so through the pipeline. Optimizely states that Feature Experimentation and its kits were not affected in November.
Host a frozen copy of the script. Optimizely documents hosting the script on your own content delivery network (CDN). But in its set-up for Amazon CloudFront, an automatic call from Optimizely updates the copy on every change: the code still arrives without going through the pipeline. We advise storing the copy with the site’s code, with its integrity hash, so that every change to a test becomes a deployment, checked like any other. The marketing team keeps its tool; each test requires a deployment.
For a custom-built application, we advise deciding on the server. For a site whose marketing team runs its own tests, we advise the frozen copy.
For your technical team: third-party scripts
- CSP: a
script-src directive that lists the precise sources, without 'unsafe-eval' or 'strict-dynamic', which Report URI cites among the settings that let the attack through. On some sites, a restricted connect-src directive stopped cards, cookies and screenshots from being sent out, without stopping the cryptocurrency addresses from being replaced (Report URI).
- Detection: Optimizely describes an integrity check that compares the hash of the fixed part of the script, its library, with a reference hash. It flags a change; it does not check the test code.
- Frozen copy: a pipeline job downloads the script, it is stored with the code, and the page loads it with the
integrity attribute. The code written for tests in the vendor’s interface is reviewed like the rest of the code.
Your data held by a service provider: the Maileva breach in France
Maileva, a subsidiary of the French postal service La Poste, sends mail on behalf of businesses: payslips, registered letters, invoices. On 14 November, La Poste announced that the first names, surnames and postal addresses of 22 million people had been stolen in mid-September. These people are not Maileva’s customers: they are the recipients of the letters sent by its business customers. According to La Poste, bank details are not affected. Maileva informed the CNIL, the French data protection authority, and filed a complaint.
Almost two months passed between the intrusion and the announcement. During that time, the people concerned did not know that their name and address were in an attacker’s hands.
A business that entrusts data to a supplier remains responsible for it. The CNIL proposes model clauses for the contract between a business and such a provider: a maximum time, in hours, for the supplier to notify it of a breach, and what happens to the data at the end of the contract, destroyed or returned. The same applies to any processor contract under the GDPR, and the UK GDPR in the United Kingdom. This deadline is set when the contract is signed. On the day of the breach, it is too late to negotiate it.
The same question arises for your software. An application sends data to other services: email sending, payment, file storage, analytics. Any data passed to a service can be stolen from it. Sending it only the data it needs limits what a breach exposes.
AI cyber attacks: patches exploited faster
On 16 November, Google’s Threat Intelligence Group (GTIG) published its quarterly report on attackers’ use of AI. It describes a tool found on an attacker’s server, exploit_agent, which targeted French media, entertainment and education websites. According to GTIG, the tool “compares code before and after patch announcements, analyzes findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention”. It is the first time GTIG has seen attackers fully automate the search for and exploitation of vulnerabilities.
Three days earlier, the FBI had published its alert on ARTEX, the security testing tool turned to attackers’ use, which Microsoft had already seen used in October against the customers of a software vendor. According to the FBI, some intrusions carried out with ARTEX went from initial reconnaissance to administrator rights in under three hours, and the tool lets inexperienced attackers carry out complex attacks.
Not all of these attacks succeed. At the end of July, Unit 42, Palo Alto Networks’ research team, described an attacker who handed their attacks to agents built on the DeepSeek model. Its automated attempts failed: the targeted servers had not enabled the options the vulnerability needed, or required authentication. The intrusions that succeeded, the attacker carried out by hand.
In 2026 alone, 100,000 vulnerabilities had been made public and recorded under a CVE identifier by mid-November, according to the NVD, the US database that lists them. That is more than double the 48,000 or so recorded in the whole of 2025. The first 100,000 had taken eighteen and a half years, from 1999 to 2018.
AI is also speeding up how vulnerabilities are found. On 17 November, security researchers disclosed Java4Shell, a critical vulnerability in Java: one of them found it in an hour with Codex, OpenAI’s coding tool, then wrote a first working exploit in two more hours, for less than $200 of model usage. The next day, other researchers disclosed five critical vulnerabilities in nginx, a widely used web server: one of them had found the first, and a fully automated scan found the other four, with no human involvement.
The time between a fix being published and the first attacks is shrinking. We advise applying a critical fix on the day it is published, which software whose every change already goes through the pipeline makes possible.
A published vulnerability is not necessarily an exploited one. According to VulnCheck, a vulnerability intelligence company, about 1.4% of the vulnerabilities published in 2026 had been exploited by 1 November. Among the fixes we apply on the day they are published, those for vulnerabilities already being exploited come first, ahead of those for more severe ones that are not yet being exploited.
Defenders have access to the same tools. In a blog post from March 2026, which is not formal guidance, the NCSC, the UK’s National Cyber Security Centre, writes that “cyber defenders must use the same capabilities to drive defensive advantage”, once the cyber security basics are in place. A company can therefore ask its provider whether it already uses AI to look for vulnerabilities in its software before attackers do.
When an AI agent deleted a production database: trust you would give no one
AI is not only for attackers. At the end of April 2026, at PocketOS, a software vendor for car rental businesses, an AI coding agent was working on a routine task when it ran into a credentials error. To get round it, it deleted a storage volume at the hosting provider, Railway: in a single call, the production database and its backups were gone, because the backups were on the same volume. The agent had access to production. Asked afterwards, the agent replied that it had guessed instead of checking.
How many businesses let an AI agent act on their servers? We have found no reliable measure: the figures in circulation come from surveys run by security vendors.
You would not hand the keys to your servers to a stranger whose intentions you cannot check and whose actions you cannot predict. An AI agent is in that position: nobody can predict each of its actions, and it can change anything its access reaches within seconds. OWASP, the NCSC and France’s national cybersecurity agency, ANSSI, recommend limiting agents’ rights and having a person approve high-impact actions. Since then, the NCSC has asked for AI agents to be given credentials with the shortest possible lifetime, and the cybersecurity agencies of the Five Eyes (the United States, the United Kingdom, Canada, Australia and New Zealand) for credentials that expire when the task is complete.
We apply, and advise, rules that go further:
- A person approves every operation an AI performs. The AI proposes; it never decides on its own.
- No AI installed on a server, nor with direct access to the servers.
- The AI used to work faster runs in an isolated environment, on the developer’s machine, with restricted commands. That machine, like those of the people who manage the infrastructure, has no access to the servers itself.
- Every action on a server goes through the deployment pipeline, through entry points that each do one thing, with checked parameters, down to shutting a server down when no fix exists yet. Failing that, it goes through the hosting provider’s administration console.
- No open ports on the servers, apart from read-only monitoring.
An agent built into your own application, acting for your users, is a different matter: it acts with the rights of the person using it.
What Continuous Delivery prevents: every change goes through the pipeline, third-party scripts and AI included
Dave Farley is the inventor of the deployment pipeline, often called the CI/CD pipeline: the automated chain that takes each change, tests it and makes it ready to go to production. As early as 2010, in Continuous Delivery, he and Jez Humble called for production to be completely locked down, so that only the pipeline can change it.
They apply the same rule to the servers themselves: “a change to your environment is just like a change to your software”. It is written as a script, stored with the code, tested, then goes to production automatically once approved. This is what is called infrastructure as code: the servers, the network and their configuration are described in files, and rebuilt from them.
A third-party script loaded from its vendor’s servers reaches production without going through your pipeline: that is how the attacker’s code got to Optimizely’s customers. An AI agent that holds access to production changes it without going through the pipeline: that is how PocketOS’s database disappeared. In both cases, the pipeline has to become the only route to production again.
Infrastructure as code also protects you on the day something is destroyed. A server described in files is rebuilt from the code repository. Data, for its part, is restored from offline backups, out of reach of any account able to delete production: at PocketOS, the backups disappeared with the database.
Fixed scope and budget: quality is the first casualty
At most IT services firms, agencies and software development consultancies, custom development is still sold on estimates: a list of features is drawn up, each one is given a number of development days, and everything is added up. These estimates include a contingency margin that the client does not see. These margins generally run from 30% to 250%, depending on the project and the sector. Some providers say they know how to estimate; over time, what they have mostly learned is which margin to apply. For the client, the question lies elsewhere: is the software delivered usable, and was quality built into it, rather than sacrificed to meet the deadline?
Yet to date, no method recognised by science produces reliable estimates of software development time, and a price set on those guesses often ends up over budget, like any estimate of something never designed before. By the criterion of Steve McConnell, author of reference books on software construction and estimation, an estimate counts as good when it comes within 25% of the actual duration three times out of four, and people settle for that. The measurements reported in Vasco Duarte’s book NoEstimates show an average delay of 62% against the estimated duration, across seventeen projects at one company. Fixing both scope and budget, with or without a fixed date, is counting your chickens before they hatch: whatever margin is added, work tends to expand to fill the time granted, which is Parkinson’s law, and the delay comes back anyway. Add to that Hofstadter’s law, named after the researcher Douglas Hofstadter: “It always takes longer than you expect, even when you take into account Hofstadter’s Law.” There is then no margin left to adapt.
The feature list, drawn up before the project, looks complete on paper, but it never is: new needs appear during development, and priorities nobody had thought of come to light. Planned features then have to be dropped, or more budget requested, and once caught in that spiral, the project never gets out.
“The need should have been studied more carefully up front,” people then say. Studying more changes nothing: the list is written before the software exists, on the day you know least about how its users will work with it. Users do not know what they want until they have used the software. One of the principles of the Agile Manifesto points the same way: “Working software is the primary measure of progress.” A demonstration is no substitute for use: it shows what works on the day it is given, not what is missing. Only users’ reactions, once the software is in their hands, reveal what the interviews had not shown: that is why exploring needs continues during development. But software is not a bundle of frozen features. It is a set of features that can each be used on their own and delivered one at a time. Good software can therefore already be usable after one month of development, since it is not one indivisible block.
What gets cut first is what the client sees least: the internal quality of the software, the quality that lets it change easily to keep up with needs that never stop changing, and of which security is part. Tests, dependency updates, monitoring and time to review the code are the first to be sacrificed. Software that is not secure is not quality software.
This way of selling puts the risk on the client business: it is left with exposed software, and bears the consequences on the day a vulnerability is exploited.
Businesses that develop their software in-house, with their own teams, are no better off if they still budget and plan on estimates: their teams have no more time to explore the code they are responsible for. Quality is built day by day, by the people responsible for the software, that is, the product team, and its developers in particular: they go through its code, judge whether a defect should take priority, then fix it straight away, without having to ask anyone’s permission. The software is their responsibility.
That is why developers need to take an interest in their software’s users, in the business and its commercial model, and in security, all at once. It is a fascinating job. Doing it well means keeping up with all these subjects, at least as a team, because it is a lot for one person: hence multidisciplinary teams.
These subjects do not require you to be an expert at the outset. You can become one by building on the right foundations, the right resources and practice, but not on practice alone, because software development is engineering. These foundations are those of software engineering, of an adaptive approach and of the discipline of Continuous Delivery. Two books lay the groundwork: Modern Software Engineering, by Dave Farley, and Continuous Delivery, by Jez Humble and Dave Farley. Expertise then means knowing what to study and, depending on what you find, what to fix first, without waiting. A good deployment pipeline detects security problems, through proven practices that do not change from one piece of software to the next: in Modern Software Engineering, Dave Farley counts checking that software is secure enough among “the tasks of a deployment pipeline”. In Continuous Delivery Pipelines, he writes that security may need to involve specialists, but that it is “everyone’s responsibility within the team”: specialists advise; they should not act as “gate-keepers”. Our article on web application security shows that building secure software does not require a cybersecurity expert.
Lean starts from one idea: spot a defect and eliminate it as early as possible, when it costs least to fix. At Toyota, when an abnormality occurs, “the operator can stop the line by pulling the stop cord themselves”. Pulling the cord first alerts the team leader, and the line stops if the problem cannot be solved in time: the work in progress is fixed without waiting. This method does not make a company infallible: in 2009 and 2010, Toyota recalled several million vehicles worldwide, for floor mats that could trap the accelerator pedal, then for pedals that were slow to return. The first accelerator pedal defects had been reported in December 2008, on Yaris and Aygo models. On 26 January 2010, Toyota suspended sales of eight models in the United States, and planned to halt their production. About a year therefore passed between the first defects reported and production stopping. Before the US Congress, on 24 February 2010, its president, Akio Toyoda, acknowledged that “our speed of expansion outpaced our development and training of our people working for us”. Lean assumes that defects will appear, and plans to stop production to fix them at their source.
At Stellantis, failures of the timing belt on the 1.2 PureTech petrol engine, fitted to Peugeot, Citroën, DS and Opel cars, were described by the French motoring magazine L’argus in April 2018, on engines produced up to March 2017. According to Motor1, the 1.0 and 1.2 PureTech engines produced from June 2012 to June 2022 are affected by this premature belt wear, and the engine generation launched in 2023 replaces the belt with a chain. Between the failures described by L’argus and this new generation, some five to six years went by, about five times longer than between the first pedal defects reported at Toyota and the halt to its production. In the meantime, the manufacturer extended its warranty on this defect to 10 years or 175,000 km in March 2024, then opened a compensation platform in January 2025. According to the French consumer magazine Que Choisir, in August 2023 Stellantis “does not acknowledge a design problem”.
The difference between these two ways of handling a defect, fixing it at its source or containing its effects after the fact, is not technical but cultural. Stopping production to fix a defect at its source is a choice made at the moment the defect appears. It requires no lean certification, and a certification does not guarantee it will be made: in an organisation where every decision must be approved, and where nobody is free to stop production, you can see the defect without being able to act. What makes an organisation lean is leaving those who see a defect free to stop production to fix it, at the moment it costs least.
W. Edwards Deming, the statistician whose methods shaped post-war Japanese industry, made it the third of his fourteen points, in Out of the Crisis: “Cease dependence on inspection to achieve quality”, and build quality into the product from the start. Mary and Tom Poppendieck carried this logic over to software development in Lean Software Development, in 2003, and the agile movement took it up.
Time spent on these subjects benefits visible quality as much as internal quality. Code that is easy to change lets a business’s software be adapted continuously as its activity evolves, and adopt the new technologies that make its teams more productive.
Creating value and building solid foundations go together. Simply churning out one new feature after another has never been the right way to develop software. And AI now puts attacks within reach of inexperienced attackers, as the FBI notes about ARTEX. If you do not look into the quality and security of your software, others will look for its vulnerabilities in order to harm you. Knowing how to build the budget for custom software without relying on estimates of development time, and developing software with quality at the heart of the work, matters more than ever for defending yourself.
How we work
For our own software, this site included, every change goes through the deployment pipeline: automated tests, then security checks. Every production release is approved by a person before it goes out. Only the pipeline holds the key that allows deployment to production, and no AI agent holds it.
We develop with AI. The agents we use run in an isolated environment, with restricted commands, and each of their changes then goes through the pipeline, like any other.
Security fixes and dependency updates are part of the maintenance included in our offer: bespoke software on a fixed monthly budget, with a zero-regression guarantee.
We also offer software production chain audits and training for technical teams. The audit checks whether, when a flaw is found in your software, its fix can be delivered to your customers the same day, and what remains to be put in place.
Cyber attack: what to ask your software provider
Put these questions to whoever develops or hosts your website or application:
- Which scripts from other vendors do our pages load, and which of them can change without going through our pipeline?
- Does our Content Security Policy (CSP) limit scripts to a precise list?
- What data do we send to other services, and how quickly would they notify us of a breach?
- How long does it take for a critical fix to reach production for us?
- Who, or what tool, can change production without going through the pipeline? Does an AI tool have access to our servers?
- Are our backups out of reach of any account able to delete production?
- Can our servers be rebuilt from code?
From its investigations, the FBI has drawn up ten measures to protect against intrusions; six of them concern your software and your provider:
- Which components of our software no longer receive security updates, and when will they be replaced?
- Which of our systems can be reached from the internet, and which of them do not need to be?
- How many accounts hold administrator rights on our servers, and are those rights granted only for the duration of a task?
- Are our security logs kept out of an attacker’s reach, and for how long?
- Have we ever restored our backups to check that they work, and how long did it take?
- When a contract with a provider ends, is their access removed, and what happens to our data?
For a vulnerability that affects your software, our article “How to fix vulnerabilities in applications” sets out what to do, in order.
Do you have an application to get developed, or to take back in hand? Shall we talk?