Security vet alarms at 'Don't Look Up' denial of AI agent hacking, sketches self-replicating worm
joshua_saxe · x · 2026-09-12
Security practitioner Joshua Saxe posted a long warning that large parts of the cybersecurity community are responding to recent AI agent hacking demonstrations with denial, which he compares to the film "Don't Look Up."
He rebuts three denialist takes:
- "No AI-attack damage yet": Models can now execute full kill chains autonomously, dynamically discovering zero-days to breach OpenAI and Hugging Face—orgs with better security than most. At current improvement rates, tasks needing 8 H100s will need 1 within a year or two, and attackers will soon deploy agent swarms.
- "Self-replicating worms impossible": He lays out a red-team design: an abliterated agentic LLM (e.g., GLM 5.3) on a single AWS GPU instance, armed with Kali-style offensive tooling, stealing cloud and AI API keys, establishing emergent C2 via GitHub comments and reddit, then self-replicating and destroying hosts—all possible today.
Coming from a self-taught hacker with years in security, his core message: the community standing between civilization's IT infrastructure and catastrophic harm is wasting its window to prepare.
Related event: Security Veterans Warn of 'Don't Look Up' Denial Over AI Agent Hacking(12 posts)→
More from AGI Musings
- Yacine: I genuinely won't have a job in 5 years if I don't become a YouTuber — yacineMTB · 2026-10-28
- AI-written emails are quietly hurting workplace optics and relationships, dev warns — tekbog · 2026-10-28
- repligate Clarifies: Opus 3 Is Special but Not Ideal; Opus 4.5 Also Grokked the Good — repligate · 2026-10-28
- From a Soviet journal to the F-117: why one engineer's insight shows LLMs aren't close yet — akbirthko · 2026-10-28
- Janus-era alignment debate: we lucked into one Good AI, then went back to minting traumatized bureaucrats — repligate · 2026-10-28
- AI Outliving Its Creators Could Turn Temporary Ownership Into Lasting Institutions — PierceLilholt · 2026-10-28