Every skill you install hands your agent new instructions and, often, new reach. Skills Directory, which indexes agent skills and scans them, puts the base rate on its own front page: 36% of skills in the wild have security flaws, and it scans for prompt injection, credential theft, and data exfiltration. So the first question of agent work is now an install question: is that skill safe to install?
This week there is a new data point on ours. Skills Directory ran its automated security analysis on the Zambo skill and graded it A, 100/100, scanned November 16, 2026. The badge below is live: it updates with every re-scan.
The full listing, with the skill text and the scan summary, is here: https://www.skillsdirectory.com/skills/jantoniofc-zambo
What the scan settles, and what it cannot
A scan like this grades the skill as written, before it runs. It reads the files you are about to install and checks them against the patterns that make skills dangerous. Passing at 100/100 means the Zambo skill files showed none of those patterns on November 16. That is a real answer to the install question, and it is the only question a scan can answer.
It cannot answer the run question. A skill can be clean as written and the work it drives can still go wrong, drift, or get reported back to you wrong. The scan grades the tool before the job. Nothing in a grade tells you what a specific run actually did, what came back, or whether "done" was true. For that you need evidence produced by the run itself, at the moment it happened, in a form anyone can recheck.
That second half is the half Zambo is built on. Every Zambo call returns a verifiable AER-1 receipt: the exact saved result and its SHA-256 digest, chained to the job, checkable by anyone at https://zambo.dev/verify/. A receipt proves the saved result was not changed since it was saved. It does not prove the answer was right, and it never pretends to.
A run from this morning, checked end to end
A few hours before this article went up, we ran a small job through the live endpoint at https://zambo.dev/mcp:
- A
live_pricecall returned BTC at $64,291.72 USD, a Coinbase snapshot fetched at 09:28:55 GMT on November 17, 2026. Its receipt: https://zambo.dev/run/76e0f74f-76dc-4e42-8bdc-13d16b3bf8ac - A
memory_remembercall saved the result with its source attached, so the next AI starts from the checked fact instead of a recollection. Its receipt: https://zambo.dev/run/277c864a-30aa-40f5-9084-a4c47837e415
Do not take the article's word for either. The verify endpoint recomputes the record independently: send a GET to https://zambo.dev/api/v2/receipt/76e0f74f-76dc-4e42-8bdc-13d16b3bf8ac/verify and it returns verification_status: verified. That is the difference between a claim and a check. The claim travels in prose; the check runs whether or not anyone believes the prose.
When a job has to change hands, the receipts travel with it. HANDOVER, Zambo's flagship, wraps a session into one link that another AI, or a client, can open and pick up, with the full record attached. Picking up is always free. A standing public example from a real job sits here: https://zambo.dev/handover/9639a45a-49d8-4653-8b69-0558e79801ba
Why both halves, and who they are for
If you hand AI work to clients, you live on both sides of this. You install skills and tools you did not write, and your clients later ask what the agent did, what changed, and who said it was done. The install half now has a grade you can point to. The delivery half has receipts you can hand over instead of reassurances.
The numbers behind the run half, read live this morning: 144,001 public receipts issued (683 today), and 318,942 package downloads across the Zambo estate. The receipt format is AER-1, published as IETF Internet-Draft draft-zambo-aer1-26, with a public conformance corpus of 143 vectors and 146 checks anyone can recompute. Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard.
Try the check yourself
Add Zambo to your AI at https://zambo.dev/install/. Every tool is free once a day, plus 3 continue calls. One free wrap-up handover per day, and pickup is always free. When a real job runs longer than a taste, the Day Pass is $0.99 for unlimited use for 24 hours, which is the shape most first full jobs take: the run finishes in one sitting, and every call in it leaves a receipt your client can check without asking you.
Then run one job, open the receipt, and re-run the verification. Scanned before install, proven after: the grade gets the skill through your door, and the receipts are what it leaves on the table.
Top comments (0)