Abstract

Attackers do not need to defeat every security control to cause serious harm. One internet-facing service, one stolen but valid credential, or one autonomous AI tool calling a legitimate interface can be enough. In November 2026, RIKEN reported that an external party reached Fugaku’s internal operational network through a publicly accessible status dashboard, tampered with internal servers and may have obtained account credentials. Separately, Google Threat Intelligence Group and Mandiant reported autonomous multi-agent offensive frameworks, including an exposed framework called exploit_agent that drafts and runs exploits without a human approving each step, and cloud-credential abuse such as LLMJacking. These are separate events, and no public evidence links them or shows AI use at Fugaku.

This paper analyzes what both events share: apparently valid access being converted into a real consequence, such as a changed configuration, a released dataset, a new privileged role, provisioned cloud resources or a dispatched payment. It explains where firewalls, authentication, IAM, zero trust, endpoint detection and AI guardrails reach their limits, without dismissing them. It then describes a defender-controlled, pre-commit enforcement point (a reference monitor, or commit gate) that holds a proposed operation in a non-effective state until its actual target, scope, destination and current authority are independently checked. The paper provides 17 diagrams, a counterfactual Fugaku analysis, AI-agent and cloud-abuse scenarios, ten falsifiable experiments, a comparison with conventional systems and 52 FAQs. It states its limits plainly. It does not claim the approach was deployed at Fugaku, would have prevented the breach, or has been benchmarked.

Based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026. The underlying invention is human-originated. AI-assisted tools were used to summarize the filing and draft the threat scenarios, comparisons and industry-relevance material, so the paper may contain errors, and readers must cross-check it against the original public sources. The company alignment (Google, Anthropic, OpenAI and Microsoft) is AI-generated, is not an endorsement, validation or infringement statement, and must be cross-checked against the original PCT filing before any conclusion is drawn. To read the full PCT filing, search “Sangam Das” on WIPO PATENTSCOPE, search PCT/IB2026/055615 (pctib202655615), or read it directly at https://www.tdcommons.org/dpubs_series/12542/. Licensed CC BY-NC 4.0.

Keywords :

AI safety; agentic AI security; AI agent security; autonomous cyberattacks; AI-enabled cyberattacks; Fugaku supercomputer breach; RIKEN; Google Threat Intelligence Group; Mandiant; exploit_agent; LLMJacking; stolen credentials; SSH key compromise; lateral movement; critical infrastructure security; supercomputer security; HPC security; cloud control plane security; zero trust; NIST SP 800-207; least privilege; reference monitor; pre-commit authorization; execution finality; non-bypassable enforcement; consequence-bound authorization; fail-safe design; cyber-physical security

Creative Commons License

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

Share

COinS