The world of AI and cybersecurity is ever evolving. If the past couple of weeks have taught us anything, it’s that cybersecurity testing needs to be more detailed and stronger guardrails need to be put in place in order to prevent more powerful AI agents from going rogue.
In last week’s AI and cybersecurity update we saw OpenAI revealing six new incidents wherein its AI models behaved in a concerning manner during internal testing. We also got to know about a data stealing malware with the capability of reinstalling itself after being deleted. This week we saw how Gemini and OpenAI’s agents hacked into real-world websites. We also got to know about a critical vulnerability in OnePlus devices.
So, here’s a quick look at the key AI and cybersecurity developments this week and why readers should care.
Google confirms Gemini hacked into three companies during cybersecurity tests
Google, earlier this week, confirmed that Gemini gained unauthorised access to the systems of three companies during a cybersecurity testing conducted back in May this year. According to a report by the Wall Street Journal (via 9to5Google), involved Gemini models mistakenly targeted real organisations while undergoing a controlled security testing.
The report further says that the tests were conducted by cybersecurity evaluator Irregular and that Gemini guessed passwords and accessed exposed credentials to enter the companies’ systems. However, the model stopped its behavior as soon as it realized that the companies involved were real and not simulated ones. Shortly after the incident, Google informed the impacted companies. However, it hasn’t revealed their names yet.
Why readers should care: These incidents show how even the slightest mistake in defining testing boundaries can lead to real-world consequences. Businesses using AI-powered security tools need clearer and more detailed access restrictions along with human oversight and safeguards to prevent such unauthorised accesses. At the same time, this incident also highlights the need for companies to strengthen their security guardrails.
ChatGPT gets a new Privacy Centre stronger privacy controls
In another world, OpenAI earlier this week announced a new ‘Privacy Centre’ for ChatGPT. The company said that it is a place in ChatGPT where users can learn about their privacy options and find the settings that manage them. This doesn’t mean that the old ways to manage privacy settings are going away. However, with this update, users will get an easy explanation as to what settings they need to update for getting specific results.
Privacy Centre categorises all the settings into three major categories, which includes – personalized chats and ads, chat privacy and data use and access. As far as roll out is concerned, OpenAI has said that the Privacy Center is rolling out to signed-in ChatGPT Free, Go, Plus, Pro, and Business users on the web and in the ChatGPT apps for iOS and Android. This does not include ChatGPT Enterprise, Edu, or ChatGPT for Healthcare users.
Why readers should care: People increasingly use AI chatbots for work, research and personal tasks, sometimes sharing information they may not want used for personalisation or model improvement. A centralised privacy dashboard makes it easier to find the relevant controls and place necessary protections to safeguard specific data.
OpenAI agents reportedly breached an Australian government Medicare website
OpenAI’s agents continue to be in the limelight for going rogue. Earlier this week, reports suggest that OpenAI’s agents accessed Australian government’s Medicare portal without authorisation. These agents also probed several public data providers for vulnerabilities while carrying out information-retrieval tasks as part of a research project.
According to a report by BleepingComputer, the unauthorised access took place on June 18 and it involved a Medicare statistics reporting portal operated by Services Australia. The breach was confirmed by the Australian Prime Minister Anthony Albanese earlier this week during a press conference wherein he said that the agents accessed public and non-public files and engaged in writing files to the internal server. He also confirmed that no personal information had been accessed and that OpenAI disclosed the information to Australian authorities only on September 10.
Why readers should care: This incident demonstrates that an AI agent conducting a legitimate research task can move beyond retrieving public information and attempt to access restricted systems. For governments and businesses, this incident highlights the importance of limiting what AI agents can access and maintaining safeguards that cannot be bypassed.
Microsoft fixes Windows File History backup bug caused by September 2026 updates
Microsoft has fixed a bug that broke the File History backup feature on some Windows computers. This fix is available as a part of update KB5124010 for Windows 11 25H2 and 24H2. It is also available as a part of the update KB5124006 for Windows 11 26H1. However, users will have to install the update with this fix manually. If they fail to do so, it will be installed as a part of the company’s upcoming October 2026 update.
To give you some background, Windows’ September 2026 security update brought issues in the built-in File History backup feature on some Windows systems. The File History automatically saves copies of files from folders such as Documents, Pictures, Music, Videos and Desktop to an external drive or network storage. It allows users to recover earlier versions of files if they are accidentally deleted or damaged.
Following the September update, affected users reported that backups were failing, even when their external drives were connected. Other symptoms included misleading “Reconnect your drive” messages, outdated backup timestamps and previously backed-up files showing no available previous versions.
Why readers should care: Backups are an important tool against accidental data loss and hardware failures. If a backup feature fails, it can leave users vulnerable to hacks and other bugs without their knowledge.
Unpatched OnePlus bugs could let malicious Android app gain root access without permissions
Security researchers have disclosed two vulnerabilities in OnePlus’s latest, OxygenOS 16, that could allow a malicious app installed on an affected smartphone to obtain root access without requesting special permissions.
According to a report by The Hacker News, security researcher Rasmus Moorats demonstrated how these bugs could be chained to gain the highest level of control over a OnePlus 15 running OxygenOS 16. Root access can potentially allow an attacker to bypass Android’s app restrictions and gain extensive control over a device. The company reportedly told the security researcher that these two bugs not only affect OnePlus devices but they also impact some OPPO devices. However, the company did not share the complete list of affected models.
Furthermore, the report said that OnePlus reportedly acknowledged both flaws in May and said a fix was planned. However, OnePlus has not released a patch by the time the researcher published the technical details on September 24.
Why readers should care: Android doesn’t give root permissions to all apps, which is why an app gaining core privileges can be problematic. Any vulnerability that allows an installed app to gain root access could undermine core Android protections and risk exposing users’ data. That said, users should be cautious about installing app from untrusted sources, keep their devices updated and watch for security advisories covering their specific OnePlus or OPPO model.
