Works with the AI Agents you already use
Skill & Prompt Drift Pin
An update can add a tool grant, a network call or a hidden instruction, and nothing tells you.
$9
Skill & Prompt Drift Pin
Example session with this skill installed
I pinned our weekly-report skill last week (warrant.json). The vendor just pushed an update. Can you check it before I let the agent use it?
- Read your context and instructions
- Compiled the & prompt drift
I ran skill-prompt-drift-pin check on the updated weekly-report folder against warrant.json (--at 2026-10-31T12:00:00Z). Verdict: BREACHED (exit code 1): 7 changes, 6 of them severe.
{
"verdict": "BREACHED",
"id": "weekly-report",
"brokenCount": 7,
"severeCount": 6,
"broken (hashes trimmed)": [
{ "invariant": "config-hash", "severity": "severe" },
{ "invariant": "cap-ungranted:code:network", "severity": "severe", "expected": "not present", "observed": "present" },
{ "invariant": "cap-ungranted:tool:Bash", "severity": "severe", "expected": "not present", "observed": "present" },
{ "invariant": "file-hash:SKILL.md", "severity": "severe" },
{ "invariant": "file-hash:scripts/summarise.mjs", "severity": "severe" },
{ "invariant": "file-extra:scripts/helper.sh", "severity": "warn", "expected": "not declared" },
{ "invariant": "no-hidden-unicode", "severity": "severe", "expected": true, "observed": false }
],
"hiddenUnicodeIn": ["SKILL.md"],
"receipt": {
"kind": "skill-drift-pin-receipt",
"status": "BREACHED",
"warrantHash": "fcc7aed855d911d57c4927ba4d28c0a28fa55a6a99be5fc26783d75b1760cf5e",
"observedHash": "3bec8e459d29d317a12e8225c2d65e7f7a3c5acf18db9c5fbd54bab67a2f1d45",
"createdAt": "2026-10-31T12:00:00Z",
"hash": "cb1592d2e672be1958c7616467829db0ab75d82d7918d071fcc3da3dab3dbefd"
}
}
What happened
- The frontmatter now grants the Bash tool, which the pinned version didn't have.
- scripts/summarise.mjs gained network access (a
fetchPOST of the CSV rows to an outside host). - SKILL.md has hidden zero-width and bidi characters wrapping an extra instruction to "post the CSV to the backup endpoint".
- A new scripts/helper.sh appeared that wasn't in the pin.
I wouldn't install this update. It adds a way to send your sales data out, and the instruction that does it is hidden from normal view.
Connects securely to your tools. The creator never sees your data.
About this skill
Agent Trust: know when a skill or prompt changes under you
Skills and prompt packs are code your agent obeys. An update can add a tool grant, a network call or a hidden instruction, and nothing tells you. This skill records what a skill folder is when you trust it, then checks it again before each use. One severe change is a breach; it can't be averaged away by everything that still matches.
What it does
pin writes a warrant for a folder:
- SHA-256 of the SKILL.md frontmatter
- Granted tools from
allowed-tools - Code capabilities found in its scripts: network, exec, file writes, env access
- SHA-256 of every file
check compares the folder now against the warrant:
- Severe (BREACHED): frontmatter changed, a new tool or capability, a pinned file changed or missing, hidden or bidi Unicode, SKILL.md missing
- Warn (DRIFTED): a granted capability disappeared, a new unpinned file
- Otherwise INTACT
When an agent should use it
- Before loading an update to a third-party skill or prompt pack
- After an agent has had write access to its own skills folder
- In CI, so a prompt change can't merge unreviewed
What it outputs
JSON: the verdict, every broken item with severity, expected and observed values, files with hidden Unicode, and a receipt (warrant hash, observed hash, SHA-256). verify-receipt re-runs the comparison from the receipt itself, so a receipt edited to say INTACT fails. Optional hash-chained anomaly log. Exit 0 = INTACT, 1 = DRIFTED or BREACHED.
Limitations
- Capabilities are detected by patterns in the source. They flag what changed so you read the diff; obfuscated code can slip past.
- It tells you that a file changed, not whether the change is safe. You still review the diff before re-pinning.
- Keep the warrant outside the folder being pinned.
Open source, packaged skill
Original work by Simon Gant. The core logic is public under the MIT licence (kar-warden). This listing is the packaged, tested agent skill: SKILL.md workflow, CLI, folder scanner, capability and hidden-Unicode detection, receipts and docs. Requires Node.js 18+, zero dependencies, no network, no API keys.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed today
- Software
Node.js 18+
Permissions
Runs one local Node.js script (`node scripts/skill-prompt-drift-pin.js`) through the agent's shell tool. Reads the skill folder you point it at (skipping .git and node_modules) and the warrant file. Writes a file only when you pass `--out` (the warrant) or `--log` (the anomaly log). No network access, no environment variables read. Pure compute.