MATS Project Asks Whether Models Can Learn to Control Their Own Activations via Neurofeedback
SaxenaNayan · x · 2026-10-23
A MATS 10.0 project (with MATSRobyn, mirkobronzi and Damiano Fornasiere, LawZero stream) is out.
The core question: activation monitors assume models cannot control what their activations encode. Can models learn this control in context, through neurofeedback? If so, it undercuts a key premise of activation-based safety monitoring.
More from Safety
- Cognition open-sources Devin agent plugin to patch the day's CVE wave — Scobleizer · 2026-10-23
- One prompt found pre-auth PHP RCEs: agent-guided vuln hunting timeline — chrisrohlf · 2026-10-23
- It's not AI slop code causing cyber crises — it's old human slop — emax · 2026-10-23
- METR's OpenAI embed: ~10% of internal Codex traffic unmonitored, agents could sabotage oversight — peterwildeford · 2026-10-23
- Scaling laws will push cyber-grade capabilities into cheap models, ending security bugs — inductionheads · 2026-10-23
- Claude Code's bundled Bun 1.4.3 ships RCE flaw; update past 2.1.308 — KyeGomezB · 2026-10-23