Definition
Plain language
The blind spot where a watchdog simply never looks.
As stated in the literature
An attack family exploiting actions a monitor is configured not to review at all — e.g. unmonitored in-workspace file edits or context-compaction summaries — rather than fooling the monitor's judgment.
Also called: coverage gaps, coverage-gap
Why it matters: A smarter reviewer does not help if the dangerous action never reaches it, so deciding what gets reviewed matters as much as how well it is reviewed.
For example, a reviewer that inspects every shell command but never inspects file edits will faithfully approve a harmless-looking command while the dangerous part is quietly written into a file it never reads.
Heard on the show
“Their own simulation shows what that's worth: for the coverage-gap attack, which uses actions the monitor never sees, success is twenty-six percent when the internals are hidden.”Episode 269 — How a Forged Transcript Got Model Weights Past a Safety Monitor