agentpassport

Cryptographically prove which human authorized which AI agent to do what — even 4 hops deep.

#ai-agents #identity #authorization #agentic-ai #mcp #security #oauth

The unsolved 2026 problem: ~80% of orgs running autonomous agents can't trace an agent's actions back to a human, and 45% still authenticate agents with shared API keys. OAuth/MCP handle one hop — but the delegation chain loses its anchor at hop 3-4. agentpassport fixes exactly that: signed, scope-narrowing delegation chains you can verify back to a human principal.

pip install cognis-agentpassport
agentpassport issue researcher --principal chris --scopes read,search,write --key K > p.json
agentpassport delegate p.json summarizer --scopes read,search --key K2 > p2.json   # subset only
agentpassport verify p2.json --keys '{"human:chris":"K","agent:researcher":"K2"}' --require write
# → valid:false, violation: required scope 'write' not held at final hop  ✅ escalation blocked

🔎 Example output

Real, reproducible output from the tool — runs offline:

$ agentpassport-emit --version
agentpassport 0.2.0

$ agentpassport-emit --help
usage: agentpassport [-h] [--version] {issue,delegate,verify} ...

Verifiable agent identity + multi-hop delegation.

positional arguments:
  {issue,delegate,verify}

options:
  -h, --help            show this help message and exit
  --version             show program's version number and exit

Blocks above are real agentpassport output — reproduce them from a clone.

Sample result format (illustrative values — run on your own data for real findings):

{
    "findings": [
        {
            "id": "1234567890",
            "title": "Suspicious Network Traffic",
            "description": "Potential malicious activity detected on network 192.168.1.100",
            "created_by": "cognis-connect",
            "created_at": "2023-02-15T14:30:00Z"
        }
    ]
}

Related MCP server: evermint-mcp

Usage — step by step

1. Install the tool:

   pip install cognis-agentpassport

2. Issue a passport for an agent, anchoring it to a human principal with an explicit scope set. --key signs it:

   agentpassport issue researcher --principal chris --scopes read,search,write --key K > p.json

3. Delegate to a child agent — scopes can only narrow (subset), never escalate:

   agentpassport delegate p.json summarizer --scopes read,search --key K2 > p2.json

4. Verify the chain back to the human. --keys is a JSON map of issuer-to-key; --require asserts a scope must be held at the final hop:

   agentpassport verify p2.json --keys '{"human:chris":"K","agent:researcher":"K2"}' --require write
   # -> valid:false, violation: required scope 'write' not held at final hop  (escalation blocked)
   echo $?   # non-zero when verification fails

5. Automate in CI / a gateway — verify the presented passport before honoring an agent action:

   - run: pip install cognis-agentpassport
   - run: agentpassport verify "$AGENT_PASSPORT" --keys "$TRUSTED_KEYS" --require write

Short-lived delegation (TTL / expiry)

Standing agent credentials are a blast-radius problem. Add --ttl <seconds> at issue or delegate time and the hop carries a signed exp; verify rejects the chain once it lapses. A child's expiry is clamped to never outlive its parent. Passports issued without a TTL never expire (fully backward-compatible with 0.1.x credentials).

agentpassport issue deploy-agent --principal release-bot --scopes deploy:staging --key K --ttl 900 > p.json
agentpassport verify p.json --keys '{"human:release-bot":"K"}'                 # valid now
agentpassport verify p.json --keys '{"human:release-bot":"K"}' --at 9999999999 # valid:false — expired

--at <unix> pins the clock for deterministic checks (CI, tests); omit it in production to use the wall clock. The verify output now also reports expires_at (earliest expiry in the chain).

Demos — real, runnable scenarios

Every passport under demos/ is a genuine HMAC-signed artifact produced by the library (regenerate with python scripts/build_demos.py). Each folder has a SCENARIO.md with where the data came from, the exact command, and how to act on the result.

Architecture

flowchart LR
  H[👤 Human principal] -->|issue scopes| A1[Agent: researcher]
  A1 -->|delegate ⊆ scopes| A2[Agent: summarizer]
  A2 -->|delegate ⊆ scopes| A3[Agent: tool-runner]
  A3 --> V{verify chain}
  V -->|walks back to| H
  V --> R[valid? · principal · violations]

Why it's different

Every hop is HMAC-signed and can only narrow scopes — escalation is detected. Verification walks the whole chain back to the human anchor, so you get the one thing OAuth/MCP can't give you today: accountable, multi-hop agent authorization.

Use it from any AI stack

MCP server (agentpassport mcp), JSON in/out for any agent runtime, drop-in for uncensored-fleet / LangChain / CrewAI delegation.

Prior art / standards

Aligned with IETF draft-klrc-aiagent-auth (AIMS), NIST agent-identity concept paper, MCP, and Mastercard Agent Pay tokenization. Production: anchor the HMAC demo in real PKI / SPIFFE.

Related

🤖 uncensored-fleet · 🛡️ guardpost · 🧰 toolguard · 🗂️ the suite

⭐ Star it — agent identity is the problem nobody's solved yet.

Interoperability

agentpassport composes with the 300+ tool Cognis suite — JSON in/out and a shared OpenAI-compatible /v1 backbone. See INTEROP.md for the suite map, composition patterns, and reference stacks.

Integrations

Forward agentpassport's findings to STIX/MISP/Sigma/Splunk/Elastic/Slack/webhooks via cognis-connect. See INTEGRATIONS.md.

License

COCL v1.0 — see LICENSE.