Topic

Guardrail bypass testing

Empirical testing of whether a stated safety or permission control actually blocks the behaviour it names, typically by enumerating variant inputs and measuring outcomes rather than reading documentation.

Current clusters

build1 publisher

Five non-literal git push spellings evaded Claude Code's committed deny rule

A single Bash(git push:*) entry in a throwaway repo's .claude/settings.json stopped 8 of 14 ways to push under Claude Code 2.1.278. The five that landed wrapped the command in git -c, quotes, sh -c, eval and a script file.

Publishers:dev.to

Reality

Evidence68
Adoption18
Hype gap−8
Incentives32
Confidence60