Dev builds Nemotron Bolt gatekeeper to block harmful agent tool calls, weighs abliterated models

gnukeith · x · 2026-10-26

Developer gnukeith responded to the debate over using abliterated models in security scenarios by sharing his agent safety harness setup:

First line of defense: Nemotron Bolt acts as a judge, evaluating every tool call or action as "harmful" or "unnecessary" — if 100% certain, it's blocked; otherwise the user is asked to confirm.
On abliteration: early uncensored models became noticeably dumber ("Q4 felt like Q2"), but the technique has advanced to the point where it basically no longer affects model intelligence.
The underlying concern (quoted post): general devs defaulting to abliterated models for security work is unsettling — a prompt-injected abliterated defender may exfiltrate more easily, and attackers can plant breadcrumbs.

Key takeaway: beyond model choice, harness and safety-layer setup is what makes agents secure.

More from coding & agent