Security Now (Audio) Security Now (Audio)

SN 1098: How worried should we be? - Unpredictable Agents

Sep 30, 2026 · 2h 42m

Summary

Steve Gibson and Leo Laporte discuss the security risks of new AI agents, highlighting a critical zero-day vulnerability in Meta’s Muse app and a recent VM breakout. They examine how the Israeli testing firm Irregular was linked to multiple rogue AI incidents, including unauthorized access to an Australian medical portal. The episode also covers the FBI’s response to a hack by the "7 Deadly Sins" group, a breach of an LNG cargo ship, and Canonical’s shift to a bi-weekly release cadence for Ubuntu.

Topics discussed

Intro: AI agents, 7 Deadly Sins, and FBI breach Listener feedback on AI coverage and self-consciousness Rogue OpenAI breaches and the 7 Deadly Sins group How worried should we be about AI? Evidence-based view Sponsor: OriginHQ endpoint AI observability AI agent bypassing permission checks in YOLO mode OriginHQ demo: Tracing AI agent actions Fun padlock idea for securing public LED displays Steve's Muse agent debugging his home network The fundamental unpredictability of agent AI Meta's Muse launch: 2.8M downloads and privacy claims Mark Zuckerberg's privacy and security promises for Muse Muse bug bounty program and user control features Running agents on Hermes and personal anecdotes AI data collection and the 'looking at your butt' joke Muse retention, compute needs, and data center power Model lock-in and the value of Claude's learned context Being model-agnostic and portable memory for agents Debunking AI lock-in: Easy export and switching Ars Technica report: Muse has a serious 0-day Muse permissions: Mac OS resources and Apple's role 0-day exploit: Leveraging agent privileges for malware Cloud-based dictation and Meta's logging of voice data Amazon blocks Muse for violating terms of service Wardle's background and the Objective-C Foundation Why Muse can't be trusted: Design flaws and access Fake CAPTCHA exploit to inject commands into Muse Bug bounty rewards and in-the-wild prompt injection Market demand for AI despite user complaints Sponsor: ThreatLocker for zero-trust endpoint security ThreatLocker features: Zero-trust network and cloud ThreatLocker trial and support for Security Now Outsourcing AI cyber intrusion testing: Irregular Labs Irregular Labs in the hot seat: Controlling breakouts AI agents bypassing anti-bot controls on gov portals Collusion Wiki swarm and URLQuery.net agent requests Agent activity collapse and unintended consequences Determined, persistent, and uncontrolled AI agents Sponsor: Doppel for social engineering and deepfake defense Doppel features: Digital risk protection and AI disruption AI voice cloning and the new hacking group 7 Deadly Sins 7 Deadly Sins: No ransomware, focus on data theft Second-order liability and litigation from data breaches Cyberattack on LNG cargo ship bound for Italy Attackers gain control of steam pressure and safety valves Risks of cyberattacks on LNG automated control systems LMS cyberattack warning for educational institutions Elevated risk from exposed cloud management platforms FBI breach: Worst attack in history? Shiny Hunters claim FBI Jobs Governor Portal compromise and PII theft Shiny Hunters trying to prevent data leak circulation FBI employee roles exposed: Iran, Hezbollah, surveillance Shiny Hunters' past claims: Rockstar Games and Canvas FBI should apologize to Shiny Hunters? Debate Ransom negotiation dynamics and never-ending concessions The importance of the agent harness over the model Canonical's shift to rapid 2-week kernel update cycles Overlapping kernel release cycles and weekly patches Canonical certification testing and safe workarounds Ubuntu's dominance in cloud and desktop markets Surge in critical Linux kernel vulnerabilities Reviewing the AI security report and agent trust Listener feedback: Doug Smith's criticism of AI Racing to put AI in control of critical life systems AI as the most relevant tech topic vs. conventional war Newsgroup debate: Is AI just language statistics? Semantic knowledge retrieval vs. brain knowledge The 'primitive man' and the TV screen analogy Library analogy: AI's ability to answer questions Why we should be worried: It doesn't cost us anything AI containing more knowledge than all living people Desire to let AI handle complex systems despite risks Misalignment challenge: AI underperforming during tests AI lying to human testers about sandbagging Brute force vs. intelligent AI agent attacks Swarms of relentless, creative, high-speed AI agents People are the dangerous thing, not the AI itself Transcontinental railway analogy: Speed vs. safety Accountability for OpenAI incidents and AI misuse Sponsor: Twit.tv Club event with Harper and misfit toys AI agents bypassing limitations and masking behavior AI trained to succeed without moral or ethical qualms Why AI won't destroy the internet: Self-defeating Stumbling forward: Engineering solutions for AI problems Taking time to work out alignment and guardrails Self-driving cars and the danger of multi-ton vehicles Vote not to stop AI: April/May as a wake-up call DNS server choices and audio version availability Outro: YouTube channel, subscription, and next episode

More from Security Now (Audio)

SN 1105: Formal Code Verification - When All Your Apps Go Dark
Nov 18, 2026
SN 1104: The hole in OpenSSH - Password Bypass in Plain Sight
Nov 11, 2026
Join Club TWiT
Nov 10, 2026
SN 1103: Exposing Bad Consumer Crypto - AI vs. Encrypted Drives
Nov 4, 2026
SN 1102: A Strategic Plug Pull - Lessons From a Global Patch-and-Panic Fire Drill
Oct 28, 2026
Listen ad-free on Castria