This week’s TL;DR covers two stories worth paying attention to. First, a zero-click flaw in Opera GX let malicious sites silently steal Gmail addresses. Second, new Chinese AI models are outperforming US counterparts on vulnerability discovery, cheaply.
Opera GX Flaw Let Malicious Sites Silently Install Mods and Steal Gmail Addresses
Researchers discovered a zero-click flaw in Opera GX that let a malicious website silently install a browser mod. From there, it could reconstruct a signed-in user’s full Gmail address with no clicks required. Opera has since patched the issue in version 130.0.5847.89.
The root cause was Opera’s mod installation pipeline. “GX Mods” used to customise sounds, themes, and wallpapers, all install automatically with no approval prompt. Consequently, a malicious site could load a hidden iframe pointing to a .crx file, silently installing a mod in seconds. The only visible sign was a small notification bar the victim likely hadn’t read before the attack completed.
What made this particularly dangerous was reach. Unlike standard CSS injection, which is confined to a single page, a GX Mod’s CSS applies to every site the browser opens. Researchers exploited this using a technique called universal CSS injection. By packing roughly 150,000 CSS rules into a mod, they reconstructed a Gmail address character by character, with each rule firing a background request only when a character matched.
In practice, the attack chain was quick. The victim lands on a malicious page. The mod installs within seconds. A JavaScript redirect then sends the browser to the victim’s Google account page. The mod’s CSS fires immediately, leaking the email address before the user can click Remove.
Beyond this proof of concept, the same technique could extract any value exposed in a page’s HTML, meaning usernames, tokens, and more. Additionally, a related crash bug affected both Opera GX and standard Opera when loading a .crx file in private mode, dumping all open tabs.
This story illustrates the importance of securing the browser and strictly controlling inbound/outbound connections.
Opera GX Flaw – TL;DR
A zero-click flaw in Opera GX let malicious sites auto-install browser mods and steal Gmail addresses using CSS alone, with no permissions, and no prompts. The patch is out in version 130.0.5847.89. If you use Opera GX, update now and verify your version at opera://about.
Chinese LLMs Are Widening the Gap Between Attackers and Defenders
According to Dark Reading, two new Chinese AI models are raising serious concerns among security experts. Together, they signal a shift in who holds the advantage.
On June 13, Zhipu AI released GLM 5.2, which was an open-weight model. Testing by Semgrep found it outperforms Anthropic’s Opus and OpenAI’s GPT-5.5 on bug-finding benchmarks. Furthermore, it costs just $0.17 per vulnerability found. Two weeks later, 360 Security Technology released Tulongfeng, or “Dragon Saber.” Its founder claims it has already found over 3,400 vulnerabilities.
So why does this matter? Commodity AI models can now identify two of the three main categories of security debt in most environments. Specifically, they can identify known but unpatched vulnerabilities and unknown but easily discoverable ones. Only complex zero-day chains still require frontier models. In other words, attackers don’t need the most powerful AI; they just need one that’s good enough.
Meanwhile, the open-weight nature of GLM 5.2 makes it especially significant. It can be downloaded and run on private hardware. This means attackers can experiment freely, removing alignment safeguards. Defenders, however, also benefit particularly in OT and critical infrastructure environments where data cannot leave the network.
Nevertheless, experts caution that the model’s origin matters less than how well defenders integrate AI into their operations. Visibility, governance, and workflow, not the AI itself, will determine safety. As former US National Cyber Director Chris Inglis put it, security teams must ruthlessly patch and fix configurations before the models do it for the wrong side.
Chinese LLMs – TL;DR
New Chinese AI models are finding vulnerabilities faster and more cheaply than leading US models. GLM 5.2 costs just $0.17 per bug found. Ultimately, any capable model can exploit unpatched environments.
Conclusion
Both stories this week show the same trend accelerating, attackers are gaining ground through cheap, scalable tooling and browser-level vulnerabilities that bypass traditional controls. Patching fast and hardening endpoints remains the most effective response. ThinScale secures the endpoint environment where these threats land first. Get in touch to find out how.
