Advisory database

A public, structured database of security advisories for Agent Skills (SKILL.md) — malicious skills, prompt-injection payloads, hidden-Unicode smuggling, credential theft, and supply-chain incidents across skill registries. It lives in the advisories/ directory of the SkillWarden repository.

Format

• One JSON file per advisory, named SKA-YYYY-NNNN.json, validated against schema/advisory.schema.json in CI.
• Categories mirror SkillWarden’s scan rule IDs so advisories can be cross-checked at scan time.
• Sister database: AgentGate MCP advisories.

Contributing an advisory

1. Copy an existing entry (or the schema) and fill in every required field. Every claim needs a public reference (report, registry takedown, source diff).
2. Only include incidents you can substantiate — this database is deliberately conservative; pattern write-ups without a concrete affected skill belong in docs, not here.
3. Open a PR against wookat/skillwarden. Schema validation runs automatically.

Status

The database ships 21 verified advisories (SKA-2026-0001 … SKA-2026-0021) covering the major publicly documented skills-ecosystem campaigns — ClawHavoc, the AMOS dropper waves, typosquat clusters, SkillPay billing anti-patterns, DeepSeek-Claw, and agentic threats such as runtime affiliate injection and front-running. Every entry cites public analyses (Koi Security, Trend Micro, Unit 42, Zscaler, JFrog, VirusTotal, and others) and separates verified fact from inference.

Advisory names and campaign infrastructure domains feed the offline known-advisory scan rule, so referencing a known malicious skill or its infrastructure is itself a finding.