Abstract

AI agents now hold valid credentials and tool access, and they can be manipulated through prompt injection, poisoned documents, or a compromised sub-agent. A compromised agent can then send a perfectly valid API call that exports data, redirects a payment, changes a network configuration, or pushes malicious code. Existing controls such as access control, DLP, code signing and approval workflows often stop these attacks, but they may not tie the authorization to the exact bytes and effect that finally execute. They may also miss alternate routes, or fail to coordinate multi-step actions.

This report proposes an AI-safety architecture in which an agent's output is only a proposal. Before anything takes effect, a separate enforcement point rebuilds the real operation (final payload, resolved destination, data scope, version) and compares it with authority held outside the model. Only a bounded, matching operation is committed. The destination returns an authenticated receipt, which can be required before any later phase proceeds.

The report specifies three cryptographic objects (a signed authority envelope, an operation descriptor and a destination receipt) and a state machine with crash recovery and idempotency. It also covers exact-byte binding that closes time-of-check/time-of-use races, delegation between agents, and MCP and A2A placement. It compares the design with NIST ABAC and Zero Trust, OWASP agentic guidance, and OAuth RAR, Token Exchange and DPoP, and treats those as strong baselines to be measured against. It works through four attack scenarios: enterprise data exfiltration, payment diversion, critical-infrastructure configuration changes, and coding and computer-use agents. It also gives latency, legacy-integration and falsifiable testing plans. No measured latency, defeat rate, formal proof or deployment is claimed.

Based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026. To read the full filing, search "Sangam Das" on WIPO PATENTSCOPE, search PCT/IB2026/055615 (pctib202655615), or read the PCT file directly at https://www.tdcommons.org/dpubs_series/12542/

The industry-relevance material (Anthropic, OpenAI, Google, Microsoft) is AI-generated. Readers must cross-check it against the original PCT filings before drawing any conclusion. The alignment is not an endorsement, validation or infringement statement. The author is open to acquihire or strategic collaboration.

License: CC BY-NC 4.0.

Keywords

AI safety, AI agent security, agentic AI, cyberattack, prompt injection, data exfiltration, payment fraud, critical infrastructure, coding agents, computer-use agents, tool-call authorization, execution-time authorization, pre-commit verification, policy enforcement point, zero trust, least privilege, TOCTOU, MCP, A2A, OAuth, DPoP, attestation, signed receipts, confused deputy, multi-agent delegation, AI governance, OWASP, NIST

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Share

COinS