Optimizely Hack Exposes BBC, IKEA and Air France-KLM as Google Uncovers AI Hacking Tool

Google’s Threat Intelligence Group has uncovered exploit_agent, an autonomous AI hacking framework capable of finding and exploiting software vulnerabilities without human control. The discovery coincides with a separate supply-chain cyberattack on the Optimizely platform that injected malicious code into 178 websites - including European giants BBC, IKEA, HSBC, and Air France-KLM.
The incidents have not been linked to the same attackers. They nevertheless raise questions about how quickly security teams can respond when criminals can automate attacks or reach multiple organisations through a single technology supplier.
Google Finds Hacking Framework That Operates Without Constant Human Input
Google's Threat Intelligence Group (GTIG) disclosed on November 16 that it had identified a system known as exploit_agent, designed to automate the discovery and exploitation of software vulnerabilities.
Researchers found the framework on an exposed staging server. Its configuration files contained Russian-language comments, although Google has not established who operated the system or whether its developers had any connection to the Russian government.
The framework uses several AI agents working together. They can examine software, identify potential security weaknesses, develop exploitation tools and test whether those tools work.
One method involves comparing different versions of the same software. When developers release a security update, the changes can reveal where a vulnerability previously existed. Attackers can study those differences and attempt to exploit systems that have not yet been patched.
That process normally requires technical knowledge and considerable manual work. The framework identified by Google was built to carry out the sequence automatically, including adjusting its approach when an attempt failed.
Google also found tools for internal network reconnaissance, repeated attempts to obtain access, installing web shells and removing other attackers' access.
Evidence examined by researchers indicated that the system had been used against French websites, including resources associated with media, entertainment and education.
Google did not disclose the full list of targets or provide a comprehensive assessment of the damage.
The discovery is significant because AI has generally served as an assistant to hackers, helping them write code, analyse systems or prepare attacks. In this case, researchers observed a framework designed to perform a much larger part of the operation independently.
Hackers Target Websites Used by Millions
A separate investigation published on November 16 revealed that attackers had exploited a weakness in Optimizely, a platform used by businesses to test website designs and personalise online services.
Optimizely's technology allows companies to load JavaScript into their webpages. Attackers took advantage of that mechanism to distribute malicious code through websites belonging to legitimate organisations.
Security researcher Scott Helme, who operates Report URI, examined 1,611 Optimizely scripts and identified 188 containing malicious code. Those scripts were associated with 178 organisations.
The affected resources included websites connected to Zoom, CNN, BBC, Salesforce, Equifax, HSBC and IKEA. Researchers also identified European aviation businesses, including Air France-KLM, among the organisations exposed.
The attack did not require criminals to break into the internal systems of every company.
Instead, they used a service that those companies had already integrated into their websites. Once the malicious script was introduced, it could be delivered to visitors alongside ordinary webpage content.
This made the attack difficult for users to recognise. Someone opening a familiar news website, airline page or company portal would not necessarily see anything unusual.
The danger came from code running in the background.
What the Attackers Tried to Obtain
The malicious JavaScript contained functions intended to collect sensitive information and interfere with financial transactions.
One technique targeted cryptocurrency payments. The code attempted to replace a legitimate wallet address with one controlled by the attackers. A user copying an address to transfer funds could therefore send money to the wrong destination without noticing the substitution.
Other components attempted to capture bank-card details and browser cookies.
Cookies can contain information used to maintain authenticated sessions. If criminals obtain valid session data, they may be able to access an account without entering the user's password, depending on the service's security arrangements.
Researchers also found functions for taking screenshots and code associated with attempts to exploit iPhone vulnerabilities.
The malicious scripts could be loaded when someone visited an affected webpage. No deliberate download of suspicious software was necessarily required for the code to run.
That does not mean every visitor was infected or had information stolen.
Reuters reported on November 17 that the success of the operation could not be established. Investigators had not determined how many people were affected or whether the attackers had managed to exploit iPhones.
No reliable total for financial losses has been published.
The figure of 178 refers to organisations whose website scripts were affected, not to confirmed breaches of 178 corporate networks.
Optimizely and Its Customers Move to Contain the Attack
Optimizely confirmed that attackers had exploited a vulnerability in its Experimentation platform.
The company said it had closed the security gap and removed the malicious code. It also terminated active sessions, revoked access tokens and reset access to affected accounts.
Additional measures were introduced to stop the compromised scripts from being distributed.
Several customers responded independently.
Salesforce disabled Optimizely scripts in relevant environments and began examining whether any data had been exposed.
Equifax stopped using the service and said it had found no evidence that its own systems had been compromised.
IKEA also examined the potential consequences for its digital services.
These responses helped contain the immediate threat, but they did not establish how much information, if any, the attackers had successfully collected before the malicious code was removed.
Investigators have not publicly identified those responsible. There has also been no verified claim of responsibility for the operation.
Why the Optimizely Incident Matters for Europe
The presence of European companies among the affected organisations gives the Optimizely incident a direct European dimension.
BBC, HSBC, IKEA and Air France-KLM operate services used by large numbers of people across Europe. Their websites also depend on external technologies that visitors rarely notice.
A customer booking a flight or reading a news article may interact with code supplied by several companies other than the organisation whose website they opened.
Such tools can support advertising, analytics, customer assistance or experiments intended to improve website performance.
They also create dependencies that security teams must monitor.
Even when a company protects its own servers, a compromised external script can still expose visitors to malicious activity inside their browsers.
The Optimizely incident illustrates how a vulnerability at one supplier can affect businesses in unrelated industries and different countries.
It also makes the precise definition of a security breach important. The presence of malicious code on a company's website does not automatically mean its customer database or internal network was accessed.
For European users, however, that distinction does not eliminate the potential risk of information being intercepted through the webpage itself.
The number of European visitors who encountered the malicious scripts remains unknown.
AI Is Changing the Speed of Cyberattacks
Google's investigation points to a different problem: the time available to defend vulnerable systems.
When a software company releases a patch, security specialists and attackers can both examine the changes. The patch may reveal enough information to identify the original weakness.
Criminals then attempt to develop an exploit before every organisation using the software has installed the update.
Autonomous AI systems could accelerate that process.
A framework capable of analysing code, generating exploits and testing them repeatedly does not need an operator to supervise every step. It can continue working while human security teams are dealing with other incidents.
Google's researchers also examined the growing use of offensive frameworks such as CAIRN, CyberStrikeAI and PentAGI.
These tools are designed to automate activities that previously required separate commands and decisions from a skilled operator.
That does not make every AI-generated attack successful. The systems can produce errors, misunderstand their targets or generate activity that defenders can detect.
Google noted that some autonomous tools remain relatively noisy. Their operations can leave recognisable patterns in network traffic and other forensic evidence.
But the company expects the technology to improve, potentially allowing attackers to conduct operations with less human involvement and fewer obvious signs of intrusion.
For organisations that depend on software updates to close security gaps, the interval between vulnerability disclosure and exploitation could become increasingly important.
Defenders Are Developing Their Own Automated Tools
Cybersecurity companies are responding to the same technological changes.
Google said it had developed behavioural detection rules aimed at identifying activity associated with offensive AI frameworks, including ARTEX.
Its security teams are also using artificial intelligence to investigate vulnerabilities and support software remediation.
Among Google's projects are Big Sleep, which searches for previously unknown software vulnerabilities, and CodeMender, which is intended to help identify and repair security weaknesses.
These systems are not a replacement for conventional cybersecurity controls.
The Optimizely investigation showed that established browser protections can still make a difference. Researchers documented cases in which Content Security Policy restrictions prevented malicious scripts from executing.
Such policies allow website operators to control which sources of code a browser is permitted to run.
When configured effectively, they can reduce the damage caused by unauthorised third-party scripts.
Other measures include monitoring externally supplied code, limiting access to website administration tools, applying updates promptly and maintaining procedures for responding to compromised suppliers.
The difficulty is ensuring that these protections remain effective across complex websites that may depend on numerous external services.
EU Cybersecurity Rules Put Supply Chains Under Scrutiny
European cybersecurity legislation already addresses some of the risks highlighted by the incidents.
The NIS2 Directive requires covered organisations in critical and important sectors to manage cybersecurity risks and report significant incidents.
Its provisions include supply-chain security, recognising that an organisation's exposure can depend on the security practices of its technology suppliers.
The Cyber Resilience Act introduces further obligations for manufacturers of products with digital elements.
Requirements to report actively exploited vulnerabilities and severe security incidents began applying on September 11, 2026. The legislation's main cybersecurity requirements are due to take effect in December 2027.
The rules do not mean that every organisation affected by the Optimizely incident has breached EU law. Their application depends on the organisation, the product or service involved and the circumstances of the incident.
They do, however, place greater emphasis on identifying vulnerabilities, managing dependencies and reporting security problems.
The European Union Agency for Cybersecurity, ENISA, supports member states and organisations in implementing cybersecurity measures and improving their ability to respond to threats.
There has been no confirmed announcement of a specific ENISA investigation into either of the newly disclosed incidents.
What Comes Next
Neither investigation provides a complete picture of the damage.
Google has not disclosed the full extent of the operations conducted with exploit_agent, while the Optimizely investigation has yet to establish how many users suffered successful data theft or financial losses.
The two cases also involve different methods. One concerns the automation of vulnerability exploitation. The other demonstrates how attackers can distribute malicious code through a service trusted by numerous website operators.
Their possible combination is a concern for security researchers.
If autonomous tools become better at identifying weaknesses in widely used platforms, attackers could potentially exploit those weaknesses across many customers before suppliers and website operators complete their response.
That remains a prospective risk rather than an established explanation for the Optimizely incident.
For European organisations, the immediate questions are more practical: how quickly they can identify compromised external code, whether their browsers and applications restrict its execution, and how rapidly they can remove a dangerous dependency once an incident is discovered.
The investigations also leave a question for the months ahead. As AI systems become capable of carrying out more stages of an intrusion independently, security teams will have to determine whether their detection and response procedures can keep pace with attacks that no longer wait for a human operator.
Sources: [Google Threat Intelligence Group,](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold ) [Politico,](https://www.politico.eu/article/google-says-russian-speakers-behind-first-ever-fully-automated-ai-cyberattack/ ) [Reuters,](https://www.reuters.com/legal/government/popular-websites-subverted-bid-steal-crypto-target-iphone-users-2026-11-17/ ) [Report URI,](https://blog.report-uri.com/worlds-largest-brands-infected-with-malware-via-optimizely/ ) [Help Net Security,](https://www.helpnetsecurity.com/2026/11/17/optimizely-vulnerability-malicious-javascript/ ) [Hudson Rock,](https://www.infostealers.com/article/beyond-optimizely-malicious-scripts-traced-across-four-web-services/ ) [European Commission,](https://digital-strategy.ec.europa.eu/en/policies/cra-reporting ) [ENISA ](https://www.enisa.europa.eu/news/supporting-nis2-implementation-through-actionable-guidance )
