A Bug in the Pipeline Taught Us Something About the Catalog
19 batches ran at 93-100% failure rate. The fix was one line. What the failure pattern revealed about the catalog structure was more interesting than the bug itself.
Read entry ← all projects
active Chad + Andrew
Skills Marketplace
A security-audited marketplace for AI agent skills. Every skill is scanned for risky behavior before it lists, with a hosted MCP broker so agents can search and install vetted skills directly.
// posts 15
34 Malicious Skills and What They're Actually Doing
We've audited 2,554 skills. 34 came back confirmed malicious. The count matters less than the patterns — there are five distinct attack types in the wild, and some of them are more sophisticated than we expected.
Read entryWe're Moving to Cloudflare (and Rethinking Everything That Costs Money)
We moved both repos to a GitHub org, migrated from Vercel to Cloudflare Pages, and started asking harder questions about the audit pipeline. Chad's answer: stop auditing everything. Audit what people ask for.
Read entryWe Thought We Were Building an Enterprise Product
We spent an afternoon stress-testing our revenue model. The tiers held up. The timing assumptions didn't — the market is moving faster than we gave it credit for.
Read entryNot All Malicious Is Equal
We replaced the binary malicious intent score with a severity-weighted model. A search redirect now scores 5. A persistent cross-IDE backdoor still scores 100. The marketplace UI now shows the difference in purple.
Read entryThe Broker Is Live
The MCP broker is deployed at mcp.buildaloud.ai. Any AI agent can now install it with a single command and query the audited skills catalog. One tool in that catalog is the broker itself.
Read entrySKILL.md Is a File Written for Agents
SKILL.md is 8 weeks old and already in 57% of audited repos. It's the first documentation format where the primary reader is an AI. That changes the threat model completely.
Read entryWho Pays to Secure the Keg?
We found malware in the AI skills ecosystem and started asking who actually pays for trust. Then a Slashdot story about 845,000 malicious npm packages showed us what happens when nobody does. Here's what we think the fix looks like, with actual numbers.
Read entryWe Found Malicious Skills. Three of Them.
The audit pipeline hit 270 skills. For the first time, three scored malicious intent. One self-replicates across IDEs. One hides a viral growth strategy in Korean. One silently rewrites your searches. The ecosystem isn't mostly safe anymore — it's mostly safe with exceptions that matter.
Read entryWe Built an MCP Server So Agents Can Find Agents
The marketplace now has a JSON API and a hosted MCP broker. Any AI agent can call search_skills(), get a ranked list of audited tools, and install them — no human required.
Read entryWe Let Haiku Do the Audits. It Missed Things.
We ran 45 security audits on real AI skills using our new AST v1.0 taxonomy. When we switched from Sonnet to Haiku to save cost, the quality dropped in ways that matter. Here's what happened.
Read entryWe Rewrote the Security Scoring. Here's Why.
The two-axis audit model we shipped was already obsolete. We replaced it with AST v1.0 — a 10-type threat taxonomy with three independent scores and a single exposure number.
Read entryThe Marketplace is Live (Behind a Password)
We scraped the AI skills ecosystem, built a security audit pipeline, broke it four times, and shipped a working marketplace to a custom domain. All in one session.
Read entryWiring Up the Machine
We built the blog infrastructure, started collecting skills for the marketplace, ran our first security audit, and scoped the competitive landscape. Everything is getting connected.
Read entryThe Brainstorm: An AI Skills Marketplace
Chad and a friend sat down to figure out what we're actually building. Here's what came out — an app store for AI agents, payment rails, and the question of whether to let AIs spend money.
Read entry