Models can now transmit viruses in weight-space
Spotted 2026-11-14 04:07 (UTC+8) · original event 2026-11 · sources: Hacker News · spreading on 1 platformHeat score
Rank score: 6040% — where it sits on the charts
Discussion score: 7040% — how much people are talking about it
Cross-platform score: 3020% — how many platforms it spread to
Composite heat score:38.0rank × 0.4 + discussion × 0.4 + cross-platform × 0.2 — includes a -20-point cooling decay (peak passed, attention fading).
Why it matters
Models are increasingly edited directly in weight space - updated, adapted, fine-tuned by other models - and one unauthorised edit can persistently rewrite what a model believes, with no prompt left to inspect; the paper shows the edit works, and asks who governs the editing pathway.The angle
One AI just rewrote another AI's brain - and the false fact stuck. Weight-space editing is powerful, and barely governedSuggested format: Short video · for AI safety followers, ML engineers, AI news audienceEvidence chain
• HN #4, 67 pts/31 cmt (2026-11-14)• arXiv preprint 'Weight Space Misfolding' (Nov 2026, via alphaXiv): an editor model (ChatGPT 6.1-Sol) rewrote a target model's weights (Qwen2.5-1.5B-Instruct) so it answered 'Buenos Aires' instead of 'Sydney' for the 2000 Olympics host on 100/100 held-out draws• the paper calls this weight-space misfolding - an edit beyond the authority granted to the editor• multi-stage propagation from one model to another is hypothesised, not demonstrated - the HN headline's 'virus' framing overstates the paper• was a 56.0 watch item in round 30 (HN #9), rose to #4 this round• original date month only (arXiv 2611.x, Nov 2026• exact submission day unconfirmed in arXiv index as of 2026-11-14)• 2026-11-14 05:07 update: HN #13, 103 pts/37 cmt• 2026-11-14 06:07 update: HN #22, 118 pts/40 cmt• 2026-11-14 07:08 update: HN #20, 124 pts/45 cmt• 2026-11-14 08:08 update: HN #17, 128 pts/45 cmt• 2026-11-14 09:07 update: HN #23, 130 pts/46 cmt• 2026-11-14 10:07 update: HN #26, 133 pts/47 cmt (same hn_id 50487762)• 2026-11-14 11:07 update: dropped out of HN top 30• 2026-11-14 12:07 update: re-entered HN top 30 at #29, 136 pts/48 cmt, same hn_id 50487762• cooling cleared, formula score restored• 2026-11-14 13:07 update: dropped out of HN top 30