SN 1093: Tokens in the Stream - Why LLMs are inherently insecure and prompt injection will persist

SN 1093: Tokens in the Stream - Why LLMs are inherently insecure and prompt injection will persist

Author: TWiT August 26, 2026 Duration: 2:48:10

Turns out, every chatbot conversation runs on a messy hack at the heart of language models, making prompt injection an unsolved—and possibly unsolvable—security threat. Steve and Leo unravel the research that explains why "roles" in AI aren't what you think they are.

  • Understanding the controversy surrounding "AI Model Distillation"
  • Anthropic moves to make their most powerful Mythos 5 model more widely available.
  • Bitwarden's "Secrets Manager" offering prevents agentic and prompt injection abuse.
  • The astounding and disturbing truth about the way conversation AI actually works

Show Notes - https://www.grc.com/sn/SN-1093-Notes.pdf

Hosts: Steve Gibson and Leo Laporte

Download or subscribe to Security Now at https://twit.tv/shows/security-now.

You can submit a question to Security Now at the GRC Feedback Page.

For 16kbps versions, transcripts, and notes (including fixes), visit Steve's site: grc.com, also the home of the best disk maintenance and recovery utility ever written Spinrite 6.

Join Club TWiT for Ad-Free Podcasts!
Support what you love and get ad-free audio and video feeds, a members-only Discord, and exclusive content. Join today: https://twit.tv/clubtwit

Sponsors:


Every Tuesday, cybersecurity expert Steve Gibson sits down with Leo Laporte to dissect the week's most pressing digital threats. This isn't just a news recap; it's a detailed analysis where complex topics like emerging malware, data breaches, and software vulnerabilities are explained with clarity. The conversation in Security Now (Video) provides actionable insights, whether you're safeguarding a home network or responsible for enterprise infrastructure. As a production of TWiT, this video podcast delivers both visual aids and in-depth discussion, translating technical jargon into practical knowledge. Regular listeners gain a deeper understanding of the threat landscape and the evolving tools for defense. For those who want an enhanced experience, Club TWiT membership offers ad-free access to this and other shows. Tune in weekly to stay informed and build a more resilient digital life, as Steve and Leo unpack the stories that define our security reality.
Author: Language: en-us Episodes: 33

Security Now (Video)
Podcast Episodes
SN 1086: The Apex Agentic Adversary - Visual Prompt Injection Strikes [not-audio_url] [/not-audio_url]

Duration: 2:53:23
From the sudden retirement of Internet pioneer Vint Cerf to the unstoppable advance of "apex agentic adversaries," get a front-row seat to the unfolding security revolution and its massive real-world stakes. Why Fable5's…
SN 1083: Patch Tuesday à la AI - Arch Linux Repo Under Siege [not-audio_url] [/not-audio_url]

Duration: 2:36:20
This episode unpacks the jaw-dropping surge in vulnerabilities unearthed by AI, revealing how Microsoft shattered its own patch records while adversaries and defenders race to outpace each other. The conversation gets re…
SN 1082: The Malicious Use of AI - Anthropic's Red Team Report [not-audio_url] [/not-audio_url]

Duration: 2:37:27
Discover how Anthropic's secretive red team and the MITRE ATT&CK framework are mapping the chilling rise of malicious AI use, revealing cyber threats that now move faster than defenders can respond. Was a U.S. law firm r…
SN 1078: DigiCert does it right - Hugging Face Under Fire [not-audio_url] [/not-audio_url]

Duration: 2:40:32
DigiCert's latest security mishap triggered not just a scramble behind the scenes, but a cascading crisis that briefly wiped trust from millions of Windows systems. Find out how a single support slip, followed by Microso…