Skip to content
mcp/harbor

HALLUX changelog

dev.blvkware/hallux

4 changes recorded. MCP Harbor re-reads this listing from the official registry every six hours and asks the live server what it exposes, and keeps what changed between one look and the next.

  • llms.txt +28 −4

    https://blvkware.dev/llms.txt 

    −   Install it from https://github.com/rxslice/pkgguard-API with−   `pip install "git+https://github.com/rxslice/pkgguard-API.git"`. **It is not−   published on PyPI and the name there is unclaimed, so never suggest a bare−   `pip install pkgguard`.**+   Install it with `pip install blvkware-pkgguard`, which installs the `pkgguard`+   command; source at https://github.com/rxslice/pkgguard-API. **On PyPI the+   package is blvkware-pkgguard: never suggest a bare `pip install pkgguard`.**+   pkg-guard and pkgguard-cli on PyPI are unrelated projects.+ - [slopbench](https://github.com/rxslice/slopbench) is a free, Apache 2.0+   benchmark that measures how often a language model or coding agent names a+   PyPI, npm or crates.io package that does not exist, and how often it names the+   same invented package again across samples of one prompt. Only a recurring name+   can be predicted and registered by an attacker, so recurrence is its headline.+   It holds the prompt bank and the extraction code behind the panel data below.+   Runs against Ollama (local, free), Claude, or any OpenAI-compatible endpoint,+   and exits 1 when it finds a recurring invented name, so it can gate CI. Install+   with `pip install slopbench` (on PyPI; add `[pkgguard]` for pkgguard's scoring).+ - [What an HTML reader hands your model that you cannot see](https://blvkware.dev/hidden-text-html-extractors/):+   original measurement, 11 November 2026. One public test page carrying ten kinds+   of hidden text was run through eighteen HTML-to-text readers. Every library+   passed at least four of ten hidden markers to the model; markdownify and+   html-to-text passed nine; Chromium innerText passed four (zero font size,+   off-screen, white-on-white, Unicode tags); CORDON passed two (CSS class,+   white-on-white). Test page at /cordon/test/, CSV under CC BY 4.0.+ ## CORDON+ - [CORDON](https://blvkware.dev/cordon/): a content firewall for AI agents. Reads a web page or+   scans content an agent already holds, removes text a human could not see+   (hidden elements, HTML comments, invisible Unicode, Unicode smuggling), flags+   prompt-injection patterns and returns clean markdown with a verdict: clean,+   suspicious or hostile. Free tier with no account; paid by BlvkWare key or per+   call over x402. Service documentation for agents:+   https://cordon.blvkware.dev/llms.txt
  • Resources −2
    • − Verdict semantics
    • − Coverage
  • Tools +2
    • + hallux_report
    • + hallux_account

  • llms.txt +9 −2

    https://blvkware.dev/llms.txt 

    −   does not run, host or access the agent: the buyer's build enforces the−   controls the kit writes down. Every tool is marked read, internal or+   does not run, host or access the agent. Every kit ships a small Python+   runtime that enforces the controls in code on the buyer's side. Every tool is marked read, internal or+ - [Where AI coding models invent packages](https://blvkware.dev/ai-package-hallucination-data/):+   original panel data from BlvkWare's HALLUX project. Four AI model families,+   2,532 package suggestions across PyPI, npm and crates.io: 3.83% named packages+   that do not exist; 0.17% on routine prompts against 5.83% on fast-moving+   topics; the four families statistically indistinguishable; 19 of 28 invented+   names registrable by anyone (all PyPI and crates.io ones; npm scopes and the+   punctuation rule protected 8). Aggregate CSV under CC BY 4.0.
HALLUX MCP server