AI Extensions, October 31: a security library needs a version, and a missing tool needs a way back

OWASP’s Advisor Skill and TokenEconomizer change what an assistant sees. Neither a security answer nor a cheaper context window settles permission to act.

A security advisor with a dated corpus

OWASP’s October 29 S.A.G.E announcement promises access to current GenAI resources through chat and an MCP interface. I also checked the separately named GenAI Security Advisor repository. It calls itself an Agent Skill, currently in Preview, with public release v1.1.0. Those names do not prove identical implementations. This digest reviews the documented Skill and TokenEconomizer; the separately advertised MCP server remains an unverified installation lead, not a recommended package.

Current does not mean continuously synchronized

The Skill README says its corpus is deliberately versioned and frozen. Upstream edits are not automatically pulled into an installation. Its manifest records status, source, copied commit, date and license, including superseded and linked-only resources. That makes a useful acceptance question possible: which exact resource revision supported this answer? A friendly answer about the newest guidance is insufficient if it cannot identify the material it actually read. I have not tested answer quality or the chat endpoint.

Read the document license as well as the code license

The repository LICENSE is Apache 2.0 for its own code and instructions. The README says corpus resources keep their original licenses, much of it CC BY-SA 4.0. The manifest also records linked-only resources with unconfirmed licensing. I would link those originals instead of treating the whole collection as Apache material. Framework crosswalks help locate related controls; they do not confer certification, legal compliance, risk acceptance or a completed security review.

An installation is a change to context

The README documents Claude Code, Codex CLI, Copilot and Gemini CLI setup paths. That is documentation, not proof that every current host version works. In a staging review, pin a repository revision, inspect included scripts and resources, and use synthetic architecture examples rather than private system diagrams. Ask for source titles, versions and uncertainty. A research assistant may help prepare a threat model; it does not replace an organization’s policies, architecture review, penetration testing or legal advice.

TokenEconomizer changes the available tools

The current plugin supports Claude Code, Codex, Copilot CLI and Pi, with experimental opencode support. Its actual LICENSE is Apache 2.0. Tool pruning differs by host: single tools, capability packages or whole MCP servers; Pi and opencode have no pruning. The README provides a propose-only mode and an explicit restore path. I would begin with proposals in an authorized disposable repository, preserve the original settings, and test a later task that needs an omitted tool. Historical non-use alone cannot establish future irrelevance.

Keep the economics and the stop control separate

TokenEconomizer reports costs including its own model calls, and says learned state stays locally. Model-assisted fact proposals and loop judgments run through the configured agent, so local state is not an offline-processing guarantee. Micro-skills are off by default; pruning and the loop guard are on. The research page describes stopping persistent failures, while the current plugin README says the plugin never stops a turn itself. A warning is not a spending cap. No package was installed, authenticated or executed for this digest.

The acceptance record I would keep

For the Advisor, record pinned revision, corpus entry, status, license and the answer’s citations. For TokenEconomizer, record host version, settings diff, proposed omissions, successful restore, billed cost and whether a human had to stop a run. Keep source-reading, installation permission and production permission as separate decisions. These are proposed checks, not verified outcomes. Popularity counts are omitted because they say little about whether either extension fits an actual team.

Sources

• https://github.com/GenAI-Security-Project/GenAI-Security-Advisor
• https://raw.githubusercontent.com/GenAI-Security-Project/GenAI-Security-Advisor/main/LICENSE
• https://raw.githubusercontent.com/GenAI-Security-Project/GenAI-Security-Advisor/main/corpus/MANIFEST.yaml
• https://www.linkedin.com/posts/owasp-top-10-for-large-language-model-applications_owasp-genaisecurity-aisecurity-activity-7521616629588279296-S-qh
• https://github.com/microsoft/TokenEconomizer
• https://raw.githubusercontent.com/microsoft/TokenEconomizer/main/LICENSE