We ran 384 adversarial probes against 48 agent skills
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
A Hrih News piece on skill123's test of 48 agent skills against 384 adversarial probes found that hidden-instruction injection succeeded only 1.6% of the time, while probes asking skills to exceed their declared scope succeeded 30% of the time.