EvilTokens Compromised 12,000 Inboxes and Let AI Choose the Next Victim
The AI did more than write convincing phishing messages. It read stolen mailboxes, mapped trust and told criminals which relationships were most valuable to abuse.
Rob Nidschelm
Tech tinkerer who spends too much time with home labs and Docker stacks. Into AI, Roman & Medieval history, and all things sci-fi. Avid traveller, reader and writer, with a passion for good food.
One OAuth Profile Can Turn BIG-IP APM Into a Remote Code Execution Target
F5 says attackers are already exploiting a critical BIG-IP APM flaw. The first job is not simply to patch, but to preserve evidence and find out whether the gateway has already been touched.
Chrome Fixed 11 Critical Flaws Before the CVE Databases Could Catch Up
Chrome 154 fixes 11 critical flaws across graphics and browser components, but teams cannot wait for downstream vulnerability databases before deploying it.
A Local Process Could Redirect Muse and Inherit Everything the Agent Could Reach
The weakness needed code already running on the Mac. From there, it could redirect the agent's input and inherit the reach the user had granted it.
Cisco Found a New Way to Hunt Malware by the Prompts It Leaves Behind
The malware may change its code, but it still has to explain its task to a model. Cisco Talos is turning that language into a detection surface.
D-Link Is Still Investigating Two Critical Router Flaws With Public Exploit Code
One flaw needs no authentication and the other needs only a low-privilege account. D-Link's advisory says both remain under investigation.
One WordPress Exploit Chain Exposed 18,000 Government Records
The attack did not end with a web shell. Researchers watched the operators turn WordPress access into targeted record theft across dozens of organisations.
The Malicious npm Release Had Valid Provenance Because the Build System Worked
Provenance answered where the package was built. It could not answer whether the source reaching the trusted workflow was honest.
A Secret Designed for Attackers Told the AI Agent to Stop and It Did
The defensive string did not need the model to refuse a harmful request. It made the attacking agent mistake hostile data for a new operator instruction.
A Perfect Copy of a Trusted Website Delivered Three Chained Zero-Days
The pages borrowed real content from trusted organisations while a hidden frame delivered the already shared exploit chain.