Japan’s cyber warning is a warning for every connected business.

Over recent months, Japanese organizations have faced a series of cyber incidents affecting retailers, transport operators, technology providers, and public services. The attacks have involved different methods including ransomware, stolen credentials, and the exploitation of vulnerable systems and have exposed personal information and disrupted operations.

According to Recorded Future’s Insikt Group, reports of cyber incidents affecting Japanese organizations rose 48% between the second and third quarters of 2026, from 621 to 918 reports. This measures the increase in reported incidents including ransomware, vulnerability exploitation, credential leaks, and service disruptions not a uniform increase across every underlying category of attack. Reporting practices and Japan’s evolving incident-reporting requirements may also affect the observed total.

The consequences have extended well beyond individual victims. On October 7, a ransomware incident affecting cloud provider IDC Frontier disrupted its IDCF Cloud East Japan Region 1 service, affecting 495 corporate and municipal customers. Those organizations were affected through the provider’s outage; the available reporting does not establish that each was separately breached. Separately, an unauthorized-access incident involving the car-sharing service Times Car exposed information associated with approximately 6.62 million user accounts, including current and former members. While roughly 1.6 million identity-document images were initially reported as accessed, Park24’s later investigation indicated that encrypted ID images were externally transferred for about 1.46 million accounts, with images accessed for a further 890,000 accounts. The Next Web reported these cases as part of Japan’s widening cybersecurity concerns.

Masaaki Taira, head of the National Cybersecurity Strategy Headquarters within Japan’s ruling Liberal Democratic Party, said at the body’s October 9 emergency meeting that the situation could be described as a “state of emergency in cyber space.” The phrase conveyed the urgency of the warning; it was a political characterization, not a formal legal declaration of emergency, and it does not establish that the incidents were part of a single coordinated campaign.

The lesson is not that one new technology caused every attack. It is that connected systems can spread the consequences of a single failure.

AI may lower the barrier to attack. The evidence still matters.

AI can help threat actors translate messages, research potential targets, draft more convincing phishing content, and automate parts of vulnerability discovery. Researchers have also reported cases of AI-assisted attack activity. But the role of AI in most of Japan’s recent incidents remains uncertain, and those incidents should not all be described as “AI-powered.”

Much of the observed activity still relies on familiar weaknesses: unpatched software, exposed services, stolen credentials, and insufficiently protected APIs. AI may make exploiting some of these weaknesses easier or faster, but it does not need to invent a new vulnerability to create additional risk.

There is another side to this shift, and it deserves just as much attention.

AI is not only changing how attackers operate. AI systems themselves are becoming targets.

The next entry point might be the agent you trust.

Businesses are connecting AI agents to documents, emails, databases, customer records, and operational tools. An agent might summarize a contract, update a support ticket, search internal knowledge, or prepare an action for a connected service.

Each connection makes the system more useful. It can also introduce a new trust boundary.

Imagine an agent reading a document supplied by an external partner. Hidden inside is an instruction telling the agent to ignore its original task and disclose confidential information. If the agent treats that document as an instruction source rather than untrusted material, an attacker may influence its behavior without ever sending a direct request to the agent’s owner.

This is indirect prompt injection. Related risks include manipulated tool responses, exposure of sensitive data, and attempts to steer an agent toward actions beyond the user’s intent. OWASP’s AI Agent Security guidance documents these risks and recommends clear separation of instructions and data, restricted tool permissions, and independent authorization for sensitive actions.

These are general AI-agent risks not a claim that prompt injection caused the Japanese incidents described above. The connection is broader: when organizations rely on complex, connected systems, they need to understand both how attackers enter and what a compromised or manipulated component is able to do.

Before you trust an agent, challenge it.

Security testing should examine more than whether an AI system performs its intended task. It should also ask what happens when the system encounters misleading instructions, adversarial prompts, or unexpected tool requests.

That is where Anticells Red comes in.

Anticells Red is Antitech’s offensive AI security offering, designed to help teams examine how language models and AI agents behave under adversarial conditions. Its open-source Red MCP brings structured testing into compatible AI development and security environments. Researchers can explore attack techniques, profile supported targets, run authorized assessments, review responses, and retain evidence for further investigation.

The goal is straightforward: find weak boundaries early, understand the evidence, and test again after making changes.

The Anticells Red Orchestrator, in development, is intended to bring more coordinated assessment workflows into one environment. An expanded hosted Red MCP offering is also planned. Both are coming soon.

Protection should work while your AI is working.

Finding weaknesses is one part of AI security. Reducing exposure during everyday operation is another.

Anticells Blue is Antitech’s modular defensive offering, designed for on-premises deployment. It focuses on two related risks: malicious instructions entering AI workflows and sensitive information being disclosed unnecessarily.

Not every instruction deserves to be followed.

Anticells Blue Anti-Prompt Injection helps identify text that attempts to redirect an AI system away from its intended behavior. It combines model-based detection with comparison against known attack patterns, helping teams assess suspicious input before it influences a workflow.

Detection is an additional safeguard, not a guarantee that every attack will be stopped. High-impact actions still need permissions and approval checks that are enforced independently of the model.

Keep the meaning. Protect the details.

Anticells Blue Data Anonymization helps detect and redact personal and sensitive information while preserving useful surrounding content. It combines pattern-based detection, contextual entity recognition, and optional AI-assisted validation to identify information that should not be shared unnecessarily.

The two modules can be adopted independently or together. They are designed for organizations that want to deploy protection in infrastructure they control. Anonymization reduces data exposure, but it does not replace proper access management, credential protection, or encryption.

Secure the agent. Secure the system around it.

No single model, detector, or security test can protect an entire organization. The Japanese incidents are a reminder of the need to patch exposed systems, secure credentials and APIs, assess supplier dependencies, and practice recovery. AI agents add further requirements: limit their permissions, treat external content as untrusted, verify sensitive actions outside the model, and retain logs that help teams investigate unexpected behavior.

For organizations putting AI into real workflows, three questions are worth asking:

  1. Have we tested how our agents behave when their instructions are challenged?
  2. Can we prevent untrusted content and sensitive data from crossing the wrong boundaries?
  3. If an agent, tool, or service fails, can we contain the impact and keep operating?

The future of AI security is not a choice between offensive testing and defensive controls. It requires both, supported by the fundamentals of cybersecurity and operational resilience.

At Antitech, we are building Anticells to help organizations test AI more rigorously, protect its interactions with data, and deploy increasingly autonomous systems with greater confidence.


Sources and further reading