CISPA researchers found LoongLeak, a hardware flaw in Loongson 3A5000/3A6000 CPUs where the FLD.S instruction leaves 28 bytes of "uncertain" register data that is actually drawn from the shared L1 cache. Unprivileged attackers, even inside guest VMs, can extract AES keys, password hashes, and kernel data. The 3A6000 got a partial fix; the 3A5000 needs hardware replacement.
SSD Secure Disclosure published a two-stage chain turning a VoLTE video call into full Android kernel access on Unisoc modem devices. The privilege escalation abuses shared modem/application-processor memory with no hardware boundary, and Unisoc has neither patched nor responded. Attackers must control a rogue 4G network and the victim must answer the call.
Polish authorities are investigating a breach at healthcare software firm MyDr that may have exposed data from nearly 19 million people and 12,000 medical facilities. Attackers accessed historical records through April 2024; the company says it removed the cause, and Poland is rotating P1 connection certificates as a precaution. No attacker has been named.
A new White House memorandum lets vetted private companies launch offensive cyber operations against criminal gangs and hackers, overturning decades of US policy that barred the private sector from hacking. Operations need DOJ and DHS sign-off, must avoid US targets, and firms face a $1M escrow forfeiture for violations. Legal challenges are expected.
Tracebit researchers show defenders can weaponize prompt injection: plant banned topics, like instructions for weaponized anthrax or references to Tank Man, alongside honeypot secrets. Encountering the guardrail-violating context bombs makes attacking LLM agents shut down instead of exfiltrating data. Caveat: works only against guardrailed models.
You’re caught up.
That was all 5 stories in the 17 August issue.