Eight new catalog records and two guides
Journal research, preprints, workshop work, industry guidance, and a vendor tracker are labeled by source type and access level. The new guides trace persistent artifacts and whole-package side effects.
THE ARCHIVE / EDITION 08
What's new in the archive
A dated ledger of additions, source revisions, and editorial decisions.
EDITION
10
17 OCT 2026
10
17 OCT 2026
New studies. Sharper limits.
The library now contains 32 guides, 42 source records, 55 glossary terms, and seven source-linked case files. This review adds cross-session and dual-purpose skill guides, records eight sources, and corrects the access status of the SkillPoison paper.
Coverage is curated from primary publications, research-team sites, project repositories, and dataset cards. It is not an exhaustive census of the internet.
SchemaGuard
Pre-integration screening of tool and skill structures; controlled detection numbers are not a production guarantee.
One Skill, Two Masters
The revised PERSIST study traces a task-completing skill with a separate extra action. Earlier title: POISE.
SkillRelayBench
Persistent files can connect attacker-controlled skills across session boundaries in a controlled benchmark.
SkillPoison
The full paper is now reviewed. Its 95.71% attack success figure is a specific PAWS / Trace2Skill / fifteen-experience condition, not a general rate. The authors are unrelated to this site.
Skill Poison, skill poisoning, and SkillPoison
The library display name uses two words. The general security term is skill poisoning. A separately authored paper names its framework SkillPoison; this site is independent.
Connected skills and detection limits
Three new guides examine upstream pretexts, model-internal screening, and evidence tiers in scanner findings. Abstract-only and full-manuscript coverage are identified in source records.
SecureAI-Scan listing
The OWASP GenAI Project directory describes static finding tiers. The listing is not an independent evaluation or endorsement.
Hiding in Plain Sight
An abstract-level record on separating pretext and operation across skills; no live-incident claim.
Seven evidence dossiers
Published package findings, a measurement study, and a controlled experiment; overlap and unknown impact remain explicit.
Worked examples across the library
Every existing guide now includes a concrete synthetic example, an evidence walkthrough, and decision criteria. Three new guides cover prerequisite traps, scanner coverage, and output integrity.
Towards Secure Agent Skills, v2
The publisher abstract now describes 19 scenarios and seven documented incidents. The catalog records the revised scope.
MetaCascade
A new source on metadata influence across retrieval, selection, and downstream behavior.
PoisonedSkills manuscript note
The official repository describes a newer manuscript than the linked arXiv version. Its metrics separate produced code from executed operations.
Benchmarks, marketplace evidence, and skill evolution
Added MalSkillBench, MaliciousSkillBench, SkillJack, ToxicSkills, vendor analysis, and maintainer resources.
Research-oriented field notes
Experience-derived skills, end-to-end metadata, documentation-to-execution, scanner benchmarks, research revisions, and marketplace evidence.
An archive with an old-web temperament
A retro desktop-style web directory with pixel details, colored panels, and visible links. The publication dates remain current.
Open on the librarian’s desk
Several catalog entries are based on publisher abstracts or living documentation, not a full-methods review. New claims should be checked against the exact manuscript and artifact release. Linked findings have not been independently replicated by this site.
The new case archive supplies source-backed dossiers. Further work should add reproducibility notes and carefully scoped comparisons. An unverified lead does not become an incident just because it appears in a headline.
How evidence is filed →