PLENA AI Self-Modification & Hyperagent Attestation

In March 2026, Meta and collaborators published research introducing hyperagents — self-referential AI agents that integrate two things into a single editable program: a task agent that solves the target problem, and a meta agent that modifies both itself and the task agent. The architectural shift is that the mechanism responsible for generating improvements is itself subject to improvement. Self-modification extends not only to the agent's task-solving code but to the procedure by which the agent modifies itself.

As of the time of this writing, hyperagents are at the research-paper stage. The framework requires sandboxed execution environments, and the authors have warned that the system may behave destructively in unsandboxed contexts. Production deployment timelines are uncertain. The path from research paper to consequential real-world deployment for similar AI architectures has historically run two to four years.

The documentation gap that will form on the same timeline as the deployment maturity is substantial. When self-modifying agents begin making consequential decisions — financial transactions, public-service administration, contract execution, medical recommendations, governance actions — the question of accountability will turn on documentation that does not yet exist. Under which version of its own code did the agent take this action? What was the agent's full state at the moment of the decision? Who authorized the agent to act, and within what scope? When did the agent last modify itself, and what triggered the modification? Who is responsible — the human principal, the deploying organization, the model developer — when a self-modifying agent makes a consequential error?

The attestation gap will not be filled by the AI labs themselves. The labs' incentives favor opacity over verifiability of self-modification events; the legal regimes that will eventually require accountability are still forming; the standardized documentation infrastructure that allows multi-stakeholder verification across labs, regulators, and harmed parties does not exist.

PLENA AI Self-Modification & Hyperagent Attestation establishes the receipt infrastructure for this category, in advance of the deployment maturity. The architecture extends PLENA's existing receipt-grammar foundation to the specific evidentiary needs of self-modifying systems: state-attestation at decision moments, modification-event receipts that anchor the agent's code version at the time of action, authorization-chain documentation, and multi-forum handover for safety, regulatory, and liability inquiries.

This is forward-positioned infrastructure. The page is published now because the right time to establish category prior-art is before the category forms, not after the first consequential-harm case reaches court. The page acknowledges its forward-positioned nature openly. It does not claim that hyperagents are deployed at scale; it claims that PLENA's architecture is the right foundation for their attestation when deployment maturity arrives.

Target institutional partners as the category matures: frontier AI labs (Anthropic, OpenAI, DeepMind, Meta AI, others) developing self-modifying systems internally; AI-safety audit firms (Apollo Research, METR, Anthropic's own safety practice, equivalents emerging); AI-governance regulators (EU AI Office, US NIST, UK AISI, equivalents); academic AI-safety research centers (CHAI Berkeley, FHI's successors, CSER Cambridge, equivalents globally); AI-liability law firms as the category emerges.

Forward-positioned framing preserved. The institutional market for hyperagent attestation will form on the deployment-maturity timeline. The institutional offering is documented here to establish the architecture; pricing and pilot terms will firm up as the category matures and the first consequential-harm cases produce demand for the infrastructure.

This product is built on the multi-decade arc of AI accountability inquiry. Where actually implemented and populated, the intended architecture replicates each artifact across multiple independent archives and anchors it cryptographically to public records that do not depend on any single jurisdiction or AI lab, and verifiable offline by inquiry parties. Receipts survive the AI lab's reorganization, the deployment company's dissolution, the regulator's policy changes, and the multi-administration arc of governmental AI accountability.

AI labs maintain internal documentation of model versions, training runs, deployment events, and safety evaluations. The labs' documentation is institution-controlled, subject to confidentiality protections, accessible during litigation only through formal discovery, and trustworthy to the extent that the lab's interests align with the inquiry — which they often do not.

PLENA AI Self-Modification & Hyperagent Attestation operates on the opposite principle: receipts held by the deploying organization, the affected party, and the regulatory body, independent of the AI lab's internal records. The deploying organization (the bank using a hyperagent for loan decisions, the hospital using one for triage, the government agency using one for benefit eligibility) retains its own receipts of what the agent did, under which version, with what authorization — independent of the lab that built the model.

• EU AI Act (Regulation 2024/1689), particularly Articles 26 (deployer obligations), 50 (transparency), and 67 (post-market monitoring)
• US NIST AI Risk Management Framework and the AI Risk Management Framework Generative AI Profile
• UK AI Safety Institute evaluation frameworks
• ISO/IEC 23894:2023 AI Risk Management
• Equivalent emerging regimes globally
• The broader AI-safety and AI-governance literature
• The PLENA white paper Beyond the Will

PLENA AI Self-Modification & Hyperagent Attestation does not certify AI safety. It does not evaluate hyperagent behavior. It does not constitute an AI audit. It does not provide AI-governance legal advice. It does not adjudicate AI-caused harm cases. It does not regulate AI labs or their deployments. It does not predict AI-system behavior. It produces externally anchored receipts of what a self-modifying agent's state was at decision moments, what modifications it made and when, under whose authorization, with what observed behavior pattern — receipts that survive institutional and corporate change and remain available to inquiry parties across the multi-decade arc of AI accountability.

Launches in PLENA's 8 live languages. The page acknowledges that AI-governance vocabulary varies across jurisdictions and provides translation notes for the major regulatory frameworks (EU AI Act, NIST AIRMF, UK AISI methodology) in each language. Contact hello@joinplena.com for translator inquiries.

• The Meta hyperagents paper (Zhang et al., March 2026) and the Darwin Gödel Machine literature (Zhang et al., 2025)
• The broader AI-safety alignment literature (Bostrom, Christiano, Hadfield-Menell, Carlsmith)
• AI-governance scholarship (Floridi, Cath, Coeckelbergh)
• EU AI Act commentary and US NIST AIRMF practitioner guidance
• The PLENA white paper Beyond the Will