Open Models Almanac
Methodspost-trainingreward modelling

specific method · filed under post-training

Adversarial-agent reward-hacking audit

An adversarial-agent audit is a post-training screening step in which a dedicated hack agent probes prepared RL task environments for leaks and exploitable weaknesses.1 It checks known exploit routes while searching for new ones, and its findings drive iterative cleanup and access restrictions until no successful exploit remains.2,3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. Adedicatedhackagentthenprobes theprepared environmentsforremaining leaks andexploitableweaknesses — MiMo V2 6 technical report
  2. The agent checks these known routes while searching for new ones — MiMo V2 6 technical report
  3. We use these findings to refine cleanup and access restrictions, then rerun the hack agent — MiMo V2 6 technical report

Also called Hack Agent, adversarial screening.

sources
4
models
3
labs adopt it
2
strongest
used

Documented in

Evidence

4 quotes from 4 sources, one card per model, each opening with its strongest.

How sources treat it
used 4

Mistral-Large-4

Mistral AIadopted1 source

We also employ an adversarial agent to check for reward-hacking vulnerabilities

useddata curationin Mistral Large 4 terminal-agent training dataMistral AI
↑