Work

Agentic AI Security for Europe: After Fugaku and Google's Autonomous Hacking Agent, Why Valid Credentials Must Not Be Enough to Change a Supercomputer, Cloud or Critical System

Year: 2026
Type: article
Abstract: Europe is building and relying on supercomputers, cloud platforms and AI agents that act at machine speed, and two public reports from November 2026 show how fast apparently valid access can become real damage. RIKEN reported that an external party entered Japan’s Fugaku supercomputer network through a publicly reachable status dashboard, tampered with internal servers and may have obtained account credentials, including SSH keys. Google Threat Intelligence Group and Mandiant reported autonomous multi-agent hacking frameworks, including an exposed framework called exploit_agent that drafts and runs exploits without a human approving each step, plus cloud-credential abuse such as LLMJacking. Google’s telemetry also suggested attention to certain French websites. These are separate events, and no public evidence links them or shows AI use at Fugaku. This paper asks what a defender can still do after a firewall, login or AI guardrail has been passed. It explains where conventional controls (zero trust, IAM, endpoint detection, AI guardrails, human approval) 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. That reduces what a stolen key, a compromised dashboard or a manipulated AI agent can actually change, release or provision. For the European context, the paper maps the approach to: EuroHPC-style supercomputing and research data; the NIS2 Directive (risk-management measures and incident handling); the EU AI Act (cybersecurity and robustness of high-risk AI systems, and human oversight); DORA (ICT risk in financial entities); the Cyber Resilience Act; and GDPR security of processing and personal-data-breach exposure, which is directly relevant because the Fugaku notice concerned possible leakage of users’ personal information. It also includes 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, has been benchmarked, or guarantees regulatory compliance. 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 regulatory and industry-relevance material, so the paper may contain errors, and readers must cross-check it against the original public sources. The regulatory mapping is AI-generated and is not legal advice. 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; AI-driven offensive security; multi-agent offensive frameworks; exploit_agent; LLMJacking; Fugaku supercomputer; RIKEN; Fugaku breach; supercomputer security; HPC security; EuroHPC; research data security; Google Threat Intelligence Group; Mandiant; GTIG AI Threat Tracker; stolen credentials; credential theft; SSH key compromise; lateral movement; privilege escalation; cloud credential abuse; cloud control plane security; service account abuse; GPU resource abuse; critical infrastructure security; operational technology security; industrial control systems; cyber-physical security; telecommunications security; financial infrastructure security; zero trust; NIST SP 800-207; least privilege; IAM; reference monitor; pre-commit authorization; execution finality; non-bypassable enforcement; consequence-bound authorization; commit gate; policy enforcement point; time-of-check to time-of-use; replay protection; revocation; fail-safe design; defense in depth; NIS2; EU AI Act; DORA; Cyber Resilience Act; GDPR security of processing; personal data breach; digital sovereignty; ENISA; incident response; secure-by-design; AI governance; trustworthy AI; tool-use security; MCP security; prompt injection
Author: Das, Sangam
Language: English
Open Access status: green
OpenAlex ID: W7236172113
FWCI:
Cites:
Cited by:
Related to: 10