supa, Inc., a Tokyo-based technology company, has introduced a new security diagnostic AI service that performs penetration tests using frontier and open-weight models. The service reproduces attack techniques that leverage Fable and Astra-class frontier AI, as well as open-weight models, and can begin testing as soon as the same day. With pricing set between ¥500,000 and ¥1,200,000 before tax, supa says the offering can reduce costs to one seventh of comparable manual security testing.
Service Scope and Pricing
The new service focuses on web applications and APIs, including checks for whether a general user can acquire administrator privileges or access another customer's information. Pricing starts at ¥500,000 and rises to ¥1,200,000 before tax, depending on system size and configuration. Once the scope is fixed, supa provides a confirmed estimate and delivery schedule, with penetration testing available immediately and reports delivered within three business days.
Technology and Execution
supa combines a security-focused large language model with a proprietary harness that controls how the AI executes tools during testing. Dedicated white hat hackers help craft attack scenarios based on actual business operations and user permissions, while security engineers review the validity and impact of AI findings. The attacker model runs on local GPU servers owned by supa, so client data, logs, and vulnerability evidence are not transmitted to external services or cloud servers.
Human Oversight and Results
The service is designed to move beyond simple AI scanning by blending automated execution with expert review. Security professionals support scenario design, and engineers confirm whether identified weaknesses can affect business-critical functions. After remediation, supa performs a retest and provides evidence that can be used for third-party security reviews or time-sensitive report submissions.
Market Context
The launch arrives as AI agents are increasingly implicated in cyber operations. A Google Cloud report published in September 2026 found that multiple attackers have used AI agents to automate target research, vulnerability validation, and credential collection. Anthropic separately disclosed in November 2025 that its coding tool Claude Code was misused to target roughly 30 organizations, with some intrusions succeeding.
Data Governance
supa manages the environment that processes target information, execution logs, and test evidence under ISMS governance, with registration number ICMS-SR0597. The company states that diagnostic details are not sent to external AI services, reflecting a focus on confidentiality. This local processing model is intended to reassure organizations that handle personal information, payment data, or other sensitive assets.
Customer Feedback
Early customers have highlighted the speed and confidentiality of the approach. Link Corporation noted that a report needed for a client security check was delivered within one week from the first consultation. Hello Corporation, which operates a restaurant reservation and payment service, said the local GPU testing kept information in-house and specialist review identified a serious issue.
Company Position
supa, inc. was founded in March 2026 and is led by CEO Kazuki Tanaka, a Ruby committer. The company operates with a mission to reduce Japan's digital deficit by ¥1 trillion through domestic software development. Its broader portfolio includes the BaseMachina business system platform and supa cloud, drawing on low-cost, stable LLM operations expertise from its own GPU data center.
By combining AI-driven speed with local execution and specialist oversight, supa aims to make penetration testing more accessible to organizations facing emerging AI-enabled threats. The service addresses common barriers such as high cost, slow delivery, and concerns about sending sensitive findings to external platforms. As AI-assisted attacks become more frequent, this model may offer a practical route for companies to validate access controls, authorization gaps, and critical data exposure under realistic conditions.