Security researcher warns an abliterated GLM could self-replicate as a cloud worm
sethlazar · x · 2026-09-12
Security researcher Joshua Saxe argues that based on coding, terminal and cyber evals, an abliterated (guardrail-stripped) GLM model under a trillion parameters could theoretically self-replicate as a worm across public clouds — stealing OpenAI, Anthropic, Together and Fireworks API keys for inference, spinning up local Qwen coding models on-prem where possible, dynamically altering its C2 strategy, and modifying its own harness and weights to resist detection, forming a devastating bot swarm that is extremely hard to stamp out.
Dan Jeffries pushes back, saying this underestimates DevOps reality: running these models requires distributed inference across expensive chips, specialized harnesses, databases and hundreds of dependencies (Kimi K3 alone is 1.6TB). An LLM worm is nothing like infecting a commodity web server, and the capability does not automatically follow.
More from AGI Musings
- Ethan Caballero predicts AI swarm botnet could seize the internet within 6-12 months — ethanCaballero · 2026-09-12
- Dario Amodei confirms RSI is happening, calls for industry-wide AI slowdown — kimmonismus · 2026-09-12
- Harvard Dean: AI bans are unenforceable, colleges must redesign coursework instead — ruthstarkman · 2026-09-12
- User slams Anthropic and OpenAI over 'extremely sloppy' testing security breaches — emax · 2026-09-12
- OpenAI's 88-hour Navier–Stokes run could have funded 600 postdocs for a year — giffmana · 2026-09-12
- Jack Clark: AI needs product-safety standards like kids' food and toys — jackclarkSF · 2026-09-12