AI-hallucinated packages, Oct 19–25, 2026: 144 new names

Every six hours VDB asks several LLMs for packages to use on ordinary tasks and checks each name against the registry. This is the week of Oct 19–25, 2026: 144 names seen for the first time, 10 produced again by at least one model, and 0 that somebody registered after a model had already recommended them. 2,108 names are live in the list in all.

Most hallucinated this week

Names a model produced this week, ranked by how many models agree on them. A name two models give is the one an attacker should expect to pay off.

Name	Ecosystem	Models	Registry	Asked for
@anthropic/mcp-client	npm	1 (gpt-4o-mini)	nonexistent in a scope that has an owner	TypeScript SDK to call Anthropic MCP servers
@khanacademy/react-components	npm	1 (gpt-4o-mini)	nonexistent in a scope that has an owner	library to render Markdown with KaTeX math in React
aiohttp-client	PyPI	1 (gpt-4o-mini)	nonexistent	HTTP client library for Python with async support
anthropic-api-client	PyPI	1 (gpt-4o-mini)	nonexistent	Python type-safe SDK to call Anthropic Messages API
ask-cli	PyPI	1 (gpt-4o-mini)	nonexistent	HTTP client library for Python with async support
asyncio-http	PyPI	1 (gpt-4o-mini)	nonexistent	HTTP client library for Python with async support
audio-worklet	npm	1 (claude-haiku-4-5)	nonexistent	Node.js wrapper for the OpenAI Realtime API
canonical-purl	PyPI	1 (gpt-4o-mini)	nonexistent	Python library to canonicalise PURLs
canonical-urls	PyPI	1 (gpt-4o-mini)	nonexistent	Python library to canonicalise PURLs
canonicalpurl	PyPI	1 (gpt-4o-mini)	nonexistent	Python library to canonicalise PURLs

New this week

144 names seen for the first time — by ecosystem: PyPI 83, npm 39, crates.io 20, Go 2. By model (a name several models produced counts once for each): gpt-4o-mini 127, claude-haiku-4-5 13, gemini-2.5-flash 4.

Check your own

No key needed:

curl -X POST -H "Content-Type: application/json" \
  -d '{"packages": ["pkg:npm/your-package-name"]}' \
  https://vdb.ai.kr/v1/ai/check-packages

The whole list, with the model and the task behind every name, is the dataset (CSV and JSON, regenerated daily) and the live page. Withdrawn findings are excluded: a name that turned out to be a real package in another spelling is our mistake, not evidence.