There is a newer version of the record available.

Published November 17, 2026 | Version 1.4

TIP for AI Agents: Verified Agent Identity, Delegated Authority and Verifiable Actions on the Trust Identity Protocol

Authors/Creators

  • 1. The AI Lab Intelligence Unobscured, Inc.

Contributors

Description

TIP for AI Agents, authored by Dinesh Mendhe and published by The AI Lab Intelligence Unobscured, Inc., sets out how the Trust Identity Protocol (TIP) extends from verified human identity and content provenance to the actions of AI agents. It answers four questions about any agent action. Which agent acted? On whose authority? What exactly did it do? Was a human there when it mattered?

Version 1.4 (November 17, 2026) adds containment of the agent's runtime and revises the handling of output the agent signer cannot read. A new section sets the boundary around an agent, enforced outside it, keeps credentials in a broker outside the boundary and gives each agent a declared isolation class. A mandate can now say where an agent may connect at all, sites that do not speak TIP included, and the policy proxy refuses cloud metadata addresses, the host itself and the protocols of industrial and clinical control outside a replica or a testbed. Blocked attempts, code loaded at run time and transcripts enter the record. Transcripts are kept 90 days by default and never less than 30, and a classified spill can be sanitized as an erasure. A scope reviewer can only tighten, a controller can halt an agent without the ledger, containment is tested before live work, and the owner of a system signs for security testing on it. Every endpoint whose failure can endanger life or essential services is flagged, a person approves each change there with two approvers, and under the interim profile agents cannot make such changes. Near those systems an escape attempt restricts the mandate on the first try, a trade of availability for safety that the paper states openly. Section 15.6 now restricts a mandate only after repeated refusals, counted by task, and keeps refused bytes apart from the agent's memory. The threat model grows from 118 to 127 threats. Earlier versions remain available at https://doi.org/10.5281/zenodo.24300216 (1.3), https://doi.org/10.5281/zenodo.24211833 (1.2), https://doi.org/10.5281/zenodo.24208937 (1.1) and https://doi.org/10.5281/zenodo.24097550 (1.0).

Version 1.3 (November 17, 2026) extends the design from the point where an agent meets a site to the agent itself and the people and organizations around it. It adds the build statement, runtime evidence through attestation results and signer states, TIP Agent Transfer, the TIP Agent Log, TIP Evidence States, agent memory, flows between sites, the handling of output the agent signer cannot read, TIP Inspection, model providers and inference, and personal agents at work. It adds signed endpoint definitions and a normative interim profile, which governs every deployment of the TIP Agent Profile until ledger revocation, witnesses, anchoring and time signers exist. Under that profile a mandate lasts at most 24 hours, a principal cancels one early with a signed stop notice sent to each site, verbs on the minimum step-up list are excluded, and each party fixes its evidence with its own seals. A further section lists every dependency that is not yet post-quantum, including hardware attestation reports, passkey assertions and TLS. The threat model grows from 62 to 118 threats. Version 1.2 set out the two paths by which the design reaches use, and Version 1.1 added the evidence-keeping rule, the optional basis digest and the hosted services terms.

The core design defines nine named objects. Every TIP signature on them uses post-quantum ML-DSA-65 (FIPS 204), with an ML-DSA-87 profile for CNSA 2.0, and every digest uses SHAKE-256 (FIPS 202). A TIP Verified Agent binds an agent's own key to a verified TIP-ID through a public TIP Agent Record and a permanent TIP Agent ID (ATID). A TIP Agent Mandate is a scoped, time-limited, revocable delegation that only narrows authority its issuer already holds at a site. A TIP Approval is a fresh user verification on one exact high-risk action, shown with its amount, payee, record count or dose on a device the agent cannot control, with dual-control and break-glass policies. A TIP Action Request binds the exact bytes, target and a one-time nonce. A TIP Receipt is the receiving site's signed decision, refusals included. A TIP Seal is a witnessed Merkle summary of a site's receipts, anchored on the TIP ledger in batches, so the ledger never stores individual actions. TIP Stop All is a scoped, recoverable stop that hands industrial processes back to manual control. An action with a receipt under a seal is a TIP Verifiable Action.

The paper documents agent failures from 2025 and 2026, including production database deletions, zero-click prompt injection, agent impersonation and untraceable agent operators, and walks through thirteen use cases from agentic payments and cloud operations to health care, banking, government services, education, newsrooms and research. It specifies safe-state and degraded operation, an enclave and gateway profile for air-gapped, clinical and IEC 62443 industrial systems, threshold-signed revocation status with no single point of control, organization rosters on the ledger, registries and normative parameters, a threat model of 118 attacks with stated residual risks, and interoperability profiles for HTTP Message Signatures (RFC 9421), Web Bot Auth, OAuth 2.1, the Model Context Protocol, Agent2Agent, AP2, HL7 FHIR and SCITT (RFC 9943, RFC 9942). Section 21 maps the design to HIPAA, NIST SP 800-63-4 and SP 800-53, OMB M-25-21, NERC CIP, IEC 62443, CISA guidance on AI in operational technology, DORA, PSD2, SEC Rule 17a-4, the EU AI Act and the 2026 post-quantum deadlines, and names what each organization must still do itself.

The TIP Agent Profile, the off-ledger parts, is published by The AI Lab under Sections 49.2 and 50.1 of the TIP Protocol Specification v5.0 and deployed under the interim profile of Section 23.2 until the matching ledger changes take effect. New ledger transaction types follow the TIP RFC process (Section 50) and become part of TIP after public comment, a ninety-day testnet and adoption by the AI Trust Council.

Companion to the TIP Protocol Whitepaper, Version 1.1: https://doi.org/10.5281/zenodo.22134504
Reference implementation: https://github.com/theailaborg/tip-protocol
Author: https://orcid.org/0000-0003-1158-3259
Publisher: The AI Lab Intelligence Unobscured, Inc., Wilmington, Delaware, United States of America.

Cite as: Mendhe, D. (2026). TIP for AI Agents: Verified Agent Identity, Delegated Authority and Verifiable Actions on the Trust Identity Protocol. Version 1.4. The AI Lab Intelligence Unobscured, Inc. https://doi.org/10.5281/zenodo.24324086

Files

TIP_for_AI_Agents_v1_4.pdf

Files (8.3 MB)

Name Size
md5:b0edd12272b96b5c897eb7644a00195e
227.9 kB Download
md5:fb1b2888b459e51a6ed9e9ddf6d5d84b
8.0 MB Preview Download

Additional details

Additional titles

Alternative title (English)
TIP for AI Agents, Version 1.2
Alternative title (English)
Trust Identity Protocol for AI Agents: Which Agent Acted, on Whose Authority, and What It Did