signed-audit-trails-recipe – skill Bảo mật cho agent AI

Signed Audit Trails for Claude Code Tool Calls

Cookbook-style walkthrough for cryptographically signed receipts on every

Claude Code tool call. This is the teaching skill. For the runtime

implementation, install the [protect-mcp](../../../protect-mcp/) plugin.

What this gives you

Every tool call (Bash, Edit, Write, WebFetch) is:

1. Evaluated against a Cedar policy before execution. If the policy denies

the call, the tool does not run.

2. Signed as an Ed25519 receipt after execution. Receipts are

JCS-canonical and verifiable offline by anyone with the public key.

An auditor, regulator, or counterparty can verify every receipt later

(Step 5). No network call, no vendor lookup, no trust in the operator.

When to use the pattern

• Regulated environments (finance, healthcare, critical infrastructure)

where you need tamper-evident evidence of agent behavior

• CI/CD pipelines where you want to prove that a policy gate held for

every automated build step

• Multi-party collaboration where a counterparty wants to verify your

agent's behavior without trusting your operator

• Compliance contexts (EU AI Act Article 12, SLSA provenance for

agent-built software) where standard logging is not sufficient

Step 1: Install the hook configuration

Install the protect-mcp plugin with /plugin install protect-mcp. Its hooks

run evaluate.sh before each tool call and sign.sh after it. Both scripts

read the hook event from stdin, because Claude Code sets no TOOL_NAME or

TOOL_INPUT variables. See

[references/hook-wiring.md](references/hook-wiring.md) for the hook

configuration and what each script passes to protect-mcp.

protect-mcp 0.7.4 does not create the signing key, and without a key the

receipts are unsigned. Create ./protect-mcp.key once. The command never

replaces an existing key:

if [ ! -e ./protect-mcp.key ]; then
  d=$(mktemp -d) && npx [email protected] init --dir "$d" && mv "$d/keys/gateway.json" ./protect-mcp.key
fi

Give auditors the publicKey value from that file. Do not commit the file,

because it also holds the private key.

Add the private key and receipt directory to .gitignore:

echo "/protect-mcp.key" >> .gitignore
echo "/receipts/" >> .gitignore

Step 2: Write a Cedar policy

Create ./protect.cedar from the example in

[references/cedar-policy.md](references/cedar-policy.md). It allows read-only

tools and a short list of Bash commands, denies shell chaining and destructive

commands, and limits writes to the project with .. segments denied.

Step 3: Use Claude Code normally

Start Claude Code. Every tool call goes through both hooks:

You: Please read the README and summarize it.

Claude: I will read README.md.
  [PreToolUse: Read ./README.md -> allow]
  [Tool: Read executes]
  [PostToolUse: receipt rcpt-a8f3c9d2 signed to ./receipts/]

... summary of README ...

A session of 20 tool calls appends 20 receipts to ./receipts/receipts.jsonl.

Step 4: Inspect a receipt

protect-mcp 0.7.4 appends each receipt as one line of

./receipts/receipts.jsonl. Print the newest one:

tail -n 1 ./receipts/receipts.jsonl | python3 -m json.tool

The receipt is a signed v2 envelope that names the tool, and it holds no

public key. See [references/receipt-format.md](references/receipt-format.md)

for a sample and the signed fields.

Step 5: Verify the receipts

Pass the publicKey value from ./protect-mcp.key to the verifier:

PUB=$(node -p 'JSON.parse(require("fs").readFileSync("./protect-mcp.key")).publicKey')
npx @veritasacta/[email protected] --replay-chain ./receipts/receipts.jsonl --key "$PUB"

Exit codes:

| Code | Meaning |

|------|---------|

| 0 | Every receipt verified |

| 1 | A receipt failed verification (tampered, wrong key, or malformed line) |

| 2 | The receipts file could not be read |

Step 6: Demonstrate tamper detection

Change the newest receipt's decision from allow to deny:

python3 -c "
import json
path = './receipts/receipts.jsonl'
lines = open(path).read().splitlines()
r = json.loads(lines[-1])
r['payload']['decision'] = 'deny'
lines[-1] = json.dumps(r)
open(path, 'w').write('\n'.join(lines) + '\n')
"

npx @veritasacta/[email protected] --replay-chain ./receipts/receipts.jsonl --key "$PUB"

The verifier exits with code 1 and reports which line failed. The

Ed25519 signature no longer matches the JCS-canonical bytes of the

tampered payload.

Restore the field and verification passes again.

How the cryptography works

Two invariants make receipts verifiable offline across any conformant

implementation:

1. JCS canonicalization (RFC 8785) before signing. Keys sorted by

UTF-16 code units, whitespace minimized, strings not normalized. Two

independent implementations produce byte-identical signing payloads for the same

receipt content.

2. Ed25519 signatures (RFC 8032) over the canonical bytes.

Deterministic, fixed-size, no nonce dependency.

protect-mcp 0.7.4 receipts carry no link to the previous receipt, so a

deleted receipt goes undetected.

For the formal wire format see

[draft-farley-acta-signed-receipts](https://datatracker.ietf.org/doc/draft-farley-acta-signed-receipts/).

Cross-implementation interop

The receipt format has four independent implementations today:

| Implementation | Language | Use case |

|----------------|----------|----------|

| [protect-mcp](https://www.npmjs.com/package/protect-mcp) | TypeScript | Claude Code, Cursor, MCP hosts |

| [protect-mcp-adk](https://pypi.org/project/protect-mcp-adk/) | Python | Google Agent Development Kit |

| [sb-runtime](https://github.com/ScopeBlind/sb-runtime) | Rust | OS-level sandbox (Landlock + seccomp) |

| APS governance hook | Python | CrewAI, LangChain |

A receipt produced by any of them verifies against

[@veritasacta/verify](https://www.npmjs.com/package/@veritasacta/verify).

The auditor does not need to trust the operator's tooling choice: the format

is the contract.

CI/CD integration

Verify receipts in CI so a tampered receipt fails the build.

[references/ci-cd.md](references/ci-cd.md) has a GitHub Actions workflow

that runs on pushes to the default branch. It installs the signing key from a

branch-limited environment, runs the agent, verifies the receipts, and uploads

them. It does not run on pull requests, because that would hand the key to

unreviewed code.

Composition with SLSA provenance for agent-built software

When Claude Code builds and releases software (running npm install,

npm build, npm publish as tool calls), the receipt chain is the

per-step build log. SLSA Provenance v1 has an extension point for this: the

byproducts field can reference the receipt chain alongside the build

attestation.

The [agent-commit build type](https://refs.arewm.com/agent-commit/v0.2)

documents the pattern using the ResourceDescriptor shape:

{
  "name": "decision-receipts",
  "digest": { "sha256": "..." },
  "uri": "oci://registry/org/build-xyz/receipts:sha256-...",
  "annotations": {
    "predicateType": "https://veritasacta.com/attestation/decision-receipt/v0.1",
    "signerRole": "supervisor-hook"
  }
}

The SLSA provenance is signed by the builder identity; the receipt

attestation is signed by the supervisor-hook identity. Two trust domains,

cross-referenced at the byproduct layer. See

[slsa-framework/slsa#1594](https://github.com/slsa-framework/slsa/issues/1594)

for the composition discussion.

Common pitfalls

Private key in version control. The generated ./protect-mcp.key must

not be committed. The examples above add it to .gitignore. If a key is

accidentally committed, rotate it immediately. Move the key and

./receipts/receipts.jsonl to an archive, then run the Step 1 command again.

Verify the archived receipts with the old public key.

Hook payload on stdin. Claude Code sets no $TOOL_NAME or $TOOL_INPUT

variables. A hook command that passes --tool "$TOOL_NAME" sends an empty

tool name, so the policy denies every call. Read the payload from stdin as the

plugin scripts do.

Receipts directory in CI. If Claude Code runs in CI, upload receipts as

an artifact at the end of the job or the receipts are lost at job end.

Policy is missing. When ./protect.cedar does not exist, evaluate.sh

prints a warning to stderr and allows the call. No call is gated until you

create the policy in Step 2.

Related in this marketplace

• [protect-mcp](../../../protect-mcp/) — the runtime hook implementation

(use this plugin in production)

• [review-agent-governance](../../../review-agent-governance/) — require

human approval before review-surface actions; composes with protect-mcp

References

• [draft-farley-acta-signed-receipts](https://datatracker.ietf.org/doc/draft-farley-acta-signed-receipts/) — IETF draft, receipt wire format
• [RFC 8032](https://datatracker.ietf.org/doc/html/rfc8032) — Ed25519
• [

… (nội dung đầy đủ có trong gói tải về)