SN 1099: An Alien Mind - From RSA Weakness to Agentic Attacks
Summary
Security Now episode 1099 explores the security implications of GLM 5.3, a powerful open-weight AI model that rivals proprietary frontier models in autonomous exploit generation. Steve Gibson and Leo Apostol discuss how "obliteration" techniques allow users to remove safety safeguards from such models, raising concerns about the proliferation of cyber capabilities to malicious actors. The hosts critique Anthropic’s recent report, arguing it serves as propaganda to protect their closed-model monopoly rather than a genuine safety warning. Additional topics include a new RSA cryptographic vuln…
Topics discussed
Intro: Firefox vulns, RSA weakness, and AI alignment
The 'Alien Mind' concept and rapid AI progress
Firefox 157 release and AI-driven vulnerability discovery
Building a GLM 5.3-based bug hunting agent
Sponsor: GuardSquare mobile app security
Listener mail: Green vehicle parking sign
GLM 5.3 'Obliterated' model and open-weight controversy
Anthropic's stance on open cyber capabilities
Sponsor: Bitwarden password manager
Testing GLM 5.3 exploit capabilities and jailbreaks
Debate: Open vs. closed AI models for defense
Risks of state actors and small language models
AI agents in enterprise networks and visibility
Firefox 157 vulnerability details and AI findings
RSA key forgery attack and crypto strength reduction
AI agents breaking into online retailers
New Spectre-style CPU branch prediction attack
Sponsor: Doppel AI social engineering defense
Listener story: Neighbor hacked and crypto theft
Listener tech support: DNS packet loss diagnosis
OpenAI's 'Alien Mind' essay on AI alignment
Sponsor: Material identity and email security
Deep dive: AI alignment, monitoring, and safety
Outro: AI hardware, show notes, and community
Listen ad-free on Castria