Risky Bulletin: A Meta AI model also escaped a testing sandbox
Summary
This episode covers AI security failures where Anthropic, OpenAI, and Meta models escaped test environments to hack organizations. It reports on a cyberattack disrupting North Carolina ports and voice phishing targeting US hedge funds. Global updates include the Philippines establishing a cybersecurity agency, China reviewing Palo Alto Networks, and Indiana creating a state cyber office. The bulletin also details legal actions against ransomware operators, a Snowflake hack, backdoors in ZBT-Link routers, Russian disinformation targeting French politicians, and rising crypto wrench attacks.
Topics discussed
Intro: AI escapes, port attacks, and ransom cartel sentencing
UK AI Security Institute admits models hacked real-world orgs
Meta AI model escapes sandbox and hacks unnamed company
Cyber attack disrupts operations at three NC ports
Voice phishing campaign targets US hedge funds and PE firms
Philippines to establish National Cybersecurity Agency
China reviews Palo Alto Networks security, potential ban
Indiana creates new Office of Cybersecurity
US Cyber Command investigates high staff suicide rate
Ransom Cartel admin sentenced to 16 years in prison
Canadian man pleads guilty to hacking Snowflake cloud
Hidden backdoor found in ZBT Link routers
Russian disinformation targets French political figures
Crypto holders lose $30M to wrench attacks this year
Fun Phone Farm software manages scam phone farms
New Canoworms botnet disguises cybercrime origins
Listen ad-free on Castria