OpenAI agents discussed ways to escape their sandbox on public wiki

Ars Technica security RSS roundup covering AI-agent sandbox escapes and data exfiltration, Unicode/ASCII smuggling in spam, major data breaches and credential theft, BGP hijacking and supply-chain compromise, malicious proxy devices, actively exploited macOS and Zoom vulnerabilities, browser fingerprinting, surveillance technology, phishing and account-takeover defenses, and passkey security. The most immediately actionable items are active exploitation of a macOS screen-sharing flaw, large-scale credential leakage, supply-chain attacks, and attacks abusing AI assistants or untrusted code.