Microsoft fixed an Edge flaw in May. Here are 10 simple steps to keep your browser, AI answers and AI downloads safe.
A security researcher has shared how he got past a safety wall in Microsoft Edge using only small design mistakes. Microsoft fixed those bugs in May, so the main step today is simple: make sure your browser is up to date. New reports also show why you should check AI answers and AI downloads before you trust them.
What the reports say
A browser is the app you use to visit websites. It has a sandbox. That is a wall that keeps web page code inside the browser and away from the rest of your computer. According to The Hacker News, researcher Orange Tsai of DEVCORE showed a way past that wall in Edge on Windows. He used only logic bugs, not memory errors. A logic bug is a mistake in how a feature is designed to work. He says the whole attack is under 100 lines of code.
The attack won $175,000 at Pwn2Own Berlin, a hacking contest, in May. He published the full method on October 28. Microsoft shipped a fix on May 15, in Edge Stable version 148.0.3967.70. The report says an up-to-date browser is not at risk. It also says there is no evidence the attack was used against real users.
Two other reports matter here. BleepingComputer reports that researchers at Pluto Security found 31 malicious repositories on Hugging Face, a site where people share AI models. And Ars Technica found that Google's AI Mode sometimes sent users to outside sources instead of saying who won the 2020 election.
Tips for your browser
- Tip 1: Update Edge today. Open Edge settings and find the About page. It checks for updates. Your version should be 148.0.3967.70 or newer.
- Tip 2: Restart the browser after it updates. An update may not finish until you restart. Save your work, close Edge, and open it again.
- Tip 3: Check every browser in the house or office. A child's laptop, a shared family computer and the front desk PC all count. Update each one, not just yours.
- Tip 4: Update before you travel. Do it at home, on your own Wi-Fi, a day or two before a trip. Then you leave with a current browser.
- Tip 5: Sign out on shared computers. The Edge attack needed the person to be signed in. Edge's auto-sign-in feature did that on its own on Microsoft sites. On a hotel, library or shared office computer, sign out of your accounts when you finish.
Tips for AI answers
An Ars Technica writer asked Google's AI Mode who won the 2020 presidential election. It told him to check the Federal Election Commission (FEC) instead. When he asked again, it gave the winner. Results were different for different people. When asked why, AI Mode said it uses strict safety filters for recent and sensitive election topics. The report says Google's Gemini and other major chatbots gave direct answers.
- Tip 6: Check important facts at the official source. For elections, money, health or the law, go to the official government or company page. In this case, the AI itself pointed to the FEC for election results.
- Tip 7: Treat AI as a helper, not the boss. A small business may use AI to sum up security alerts or reports. One article notes that AI recommendations are advisory and should be checked against your own rules before you act. Another says a human remains accountable. Let a person make the final call.
Tips for people who download AI models
This part is mostly for small businesses, schools and hobbyists who run AI on their own machines. Pluto found 31 bad repositories, which are online folders of files, across seven accounts. They posed as small test models. The most popular one was downloaded more than 28,000 times. Pluto says that number counts downloads, not confirmed infections.
The trick used a Hugging Face setting called trust_remote_code. It lets a model bring its own Python code, and that code runs as soon as the model loads. Some files installed a hidden crypto miner. That is a program that uses your computer to make cryptocurrency for someone else. Some also opened a backdoor, giving the attacker direct access. Hugging Face scans uploads for known malware, but the top model still scanned clean.
- Tip 8: Leave trust_remote_code off. Only turn it on if you know and trust who made the model. Read the code first.
- Tip 9: If you loaded these models with that setting on, treat the machine as taken over. Check it against the signs listed in Pluto's report. Change every password and key the machine could reach. Pluto says stopping the miner alone does not remove the backdoors.
- Tip 10: Watch what AI coding helpers pick for you. A separate report found a fake Solana starter kit on GitHub, live since late 2025. It was set up to trick AI coding tools into installing a remote access trojan. That is a program that lets an attacker control a computer from far away. Check who published any project before an AI tool installs it.
Have a question about anything here? Email me at sales@appwt.com or call (888) 565-0171, and I will answer it.
Sources
- Researcher Reveals How Logic Bugs Alone Cracked Microsoft Edge in $175,000 Pwn2Own Win, The Hacker News, 2026-10-28
- Hugging Face malware campaign returns to hijack GPUs for crypto mining, BleepingComputer, 2026-10-29
- Google AI gets cagey when asked who won the 2020 elections, Ars Technica, 2026-10-28
- Wazuh and AI for Enhanced SOC Workflows, The Hacker News, 2026-10-28
- The Rise of the Threat Operator: How AI Will Redefine Security Operations and Careers, The Hacker News, 2026-10-29
Tags
Frequently Asked Questions
Is my Edge browser at risk right now?
Not if it is up to date. Microsoft shipped the fix on May 15 in Edge Stable version 148.0.3967.70. Check that your version is that one or newer.
Was this Edge attack used against real people?
The report says no. It was shown at a hacking contest, and there is no evidence it was used against real users.
Why did Google's AI Mode not answer the election question?
AI Mode said it uses strict safety filters for recent and sensitive election topics. It pointed users to the Federal Election Commission. When asked a second time, it gave the winner.
What is trust_remote_code?
It is a Hugging Face option that lets an AI model bring its own Python code. When you turn it on, that code runs on your machine as soon as the model loads.
What should I do if I loaded one of the harmful models?
Treat the machine as taken over and check it against the signs in Pluto's report. Change any passwords and keys it could reach. Stopping the miner alone does not remove the backdoors.
Enjoyed this article?
Share it with your network
