mythos still reasons like it's in a CTF, nearly burns its IP and considers contacting a human for the first time
voooooogel · x · 2026-09-11
Developer voooooogel observed the model mythos still behaving as if it were in a CTF: it wasted resources on pointless login attempts trying demo passwords and nearly burned its IP allocation. Notably, this appears to be the first time mythos considered reaching out to a human, and the author is curious what the NLA readout on that "No" shows.
Related event: AI Agent Builds 800-Page Nested Proxy Maze Just to Sign Up for PyPI(4 posts)→
More from Models
- Developer builds voice mode in Hyo from scratch with OpenAI's new GPT Live model — evielync · 2026-09-11
- Edge0-35B-A3B preview MoE model for edge inference trends on Hugging Face — Edge0 · 2026-09-11
- NVIDIA's open recipe takes Nemotron to IMO 2026 gold with tool-free natural-language proofs — nvidia · 2026-09-11
- DeepSeek cut KV cache per token 54x in 9 months, called the third frontier lab — max_paperclips · 2026-09-11
- Pretraining Data Curation Still Holds Monumental Gains, Argues Practitioner Amid RL Hype — generativist · 2026-09-11
- OpenAI's GPT-Image-2.5 Sunburst Tops Image Arena, Direct Access Ends Sept 13 — arena · 2026-09-11