▣ Skill Poison — Knowledge ExplorerONLINE / EDITION 10
AGENT SKILLS / SECURITY
Independent. Open. Curious.
THE ARCHIVE / EDITION 08

At the evidence desk

A guide to keeping observed incidents, experiments, vendor findings, and editorial advice distinct.

This desk groups kinds of evidence. It does not claim that every linked experiment is an incident observed in a deployed system.
EVIDENCE FILE / S12

Observed marketplace behavior

A measurement study with behaviorally confirmed malicious cases. Read the collection and verification method; do not infer live package availability from an old snapshot.

Read the catalog record →
EVIDENCE FILE / S17

Vendor investigation

A vendor audit connects scanning with human-confirmed malicious findings. A flagged issue and a malicious payload are distinct labels.

Read the catalog record →
EVIDENCE FILE / S14

Controlled experiment

Adversarial skills are used to examine a documentation-to-action mechanism. The evaluation begins after loading; produced code and executed actions remain different outcomes.

Read the catalog record →
EVIDENCE FILE / S16

Benchmark construction

A curated test collection probes detection and cross-source generalization. Its class proportions do not describe the public skill ecosystem.

Read the catalog record →
EVIDENCE FILE / S13

Architecture and synthesis

A revised taxonomy maps threats and documented incidents. An architectural possibility and a demonstrated case carry different evidence.

Read the catalog record →

A compact claim ledger

RecordWhat to preserve
SourceStable identifier, version, author, original URL, check date.
MethodCollection, sample, model, host, permissions, scoring.
ObservationSuspicious text, attempted action, blocked action, or verified effect.
InterpretationWhat follows from the evidence and what remains untested.

Read the version-literacy guide →

Search guides, research, and glossary termsESC TO CLOSE

Start typing to explore the knowledge base.