Cognisant LLC: G7 Hiroshima AI Process (HAIP) Transparency Report

We do not develop foundation models and do not use numeric capability thresholds. We decide case by case. We do not build or use AI in ways that create discrimination or unfair bias, could harm life, body or property, or break the law (AI Use Policy, Art. 6), and we decline such uses. For systems that support important decisions, a person makes the final decision by design (Art. 1). For generative AI, the main risks we track are inaccurate output reaching customer work without review, and customer data being sent beyond the agreed scope. For AI agents, we treat actions that reach outside the company or cannot be undone as the main risk.

For customer systems, the customer decides on production use, based on criteria and procedures agreed with us (AI Use Policy, Art. 3). On our side, the Representative Member (the managing member and chief executive of the LLC) is accountable. Evidence: accuracy measured on evaluation questions drawn from the customer's actual work, including paraphrased questions; checks of which data fields are sent to a model; access control, logging and security settings agreed at the start of the project. Caveats: results from synthetic data or small samples are treated as an initial diagnosis and measured again on real use. Safeguards: uncertain results go to a person; we release in stages where possible and keep measuring after release. We do not currently use independent external testing.

We record incidents and security events in an internal log with the cause, the handling and the prevention measure. Staff report to the Representative Member within one hour, and prevention measures are decided within two weeks. We follow external information through a regular internal digest and public security advisories. When an issue applies to us, we update our procedures and report to any affected customer.

Anyone can write to info@cognisant.io (listed in our privacy policy, https://cognisant.io/privacy ). Customers also use the contact points set in each contract. We do not run a dedicated AI-incident form. Reports to authorities follow legal requirements, including Japan's Act on the Protection of Personal Information.

• Information-sharing partnerships (with confidentiality agreements as appropriate)
• Reporting to regulatory authorities as relevant

To affected customers: the facts of an incident, its scope, the handling and the prevention measure. To Japan's Personal Information Protection Commission: breach reports where required.

• ISO/IEC standards (e.g., ISO/IEC 27001, ISO/IEC 42001, ISO/IEC 23894, ISO/IEC 42005)
• Other international or regional standards (please specify): Japan's AI Guidelines for Business Ver1.2

On each project we work with the customer on evaluation criteria, on which data may be sent to a model, and on which results must go to a person. We do not take part in multi-stakeholder initiatives at present.

The Representative Member owns AI risk and decides on escalations. We follow our published AI Use Policy and internal security rules. As part of our ISO/IEC 27001 preparation, we keep a risk register with AI entries, and in November 2026 we started a monthly internal check of devices, account permissions and incident records. We review policies and rules every April and when laws, customer requirements or the services we use change. In November 2026 we completed a self-assessment against Appendix 7 of Japan's AI Guidelines for Business. For generative AI, the focus is on accuracy and data scope. For AI agents, it is on permitted actions and approval.

We do not train foundation models. Source documents and evaluation data come from the customer's work, and we agree with the customer how they are kept current. Personal data that is not needed is removed or masked before it is sent to a model. Formal bias testing is not yet a standard step; we are adding a check for bias factors to our project-start checklist, and we will apply it to uses that affect individuals or groups.

The Representative Member.

• Not applicable

• Other(s) (specify): staff review output before it is used in customer work (AI Use Policy, Arts. 1 and 4). Concerns can be sent to our general contact address (info@cognisant.io); a notice for this is being added to the AI Use Policy

We follow Japan's Act on the Protection of Personal Information and publish a privacy policy. Customer information is entered into generative AI only within the scope the contract allows. We use only services whose terms or settings keep inputs out of training, and we send only the fields the work needs. On some projects, the system itself limits which fields can be sent. Staff must not put work information into personal AI accounts.

• Maintain or update security policies aligned with recognized standards (e.g., NIST CSF, ISO/IEC 27001, SOC 2, CIS Critical Security Controls), and review them regularly in response to emerging risks or regulatory guidance
• Maintain vulnerability management processes (scanning, channels for external reporting, patching, remediation)
• Implement incident response, disaster recovery, or business continuity plans, including escalation procedures for high-severity incidents
• Integrate cybersecurity AI risks into broader cybersecurity risk management activities
• Regularly review security measures and update them in response to emerging risks, incidents, or major releases
• Other(s) (specify): policies and rules are explained to staff on joining, with acknowledgement recorded since November 2026; training focused on AI-specific threats is in preparation

• Apply multi-factor authentication (MFA) and role-based access controls for advanced AI systems
• Encrypt AI assets (datasets, models, configurations) in storage and in transit as appropriate
• Monitor system activity and log access to detect anomalies or misuse
• Conduct secure AI development and/or deployment practices
• Conduct security testing focused on AI-specific risks (e.g., red-teaming, adversarial testing, stress tests)
• Other(s) (please specify): physical security of the servers that host models is the responsibility of our cloud and AI model providers

• Limit access to proprietary AI assets based on role or need-to-know
• Store model weights, datasets, and algorithms in secure, access-controlled environments
• Require NDAs or confidentiality agreements for staff and contractors with access

• Monitor for unauthorized internal access or suspicious activity
• Other (please specify): see Q11A on staff training

We have declared two stars under SECURITY ACTION, the self-declaration program of Japan's Information-technology Promotion Agency (a self-declaration, not a certification). Work devices must have storage encryption, a screen lock within five minutes, automatic updates and malware protection, and the Representative Member checks the device list. MFA is required for administrator accounts and is being extended to all accounts. Our rules require a password vault. A check for secrets before commit is in place in the Representative Member's development environment and is being extended to all staff. We are preparing an ISO/IEC 27001 information security management system; we are not certified.

• Constraints on the agent's action space (e.g., tool-use permissions, environment sandboxing, allowlists/blocklists of resources, scope limits per task)
• Human oversight or interruptibility mechanisms (approval requirements for sensitive actions, escalation paths, interruption interfaces)
• Logging of agent actions
• Other (specify): documents sent outside the company go through machine checks and a separate AI review before a person approves them

We share evaluation results and their limits with the customer concerned, under the confidentiality terms of the contract. We do not publish project results that would identify a customer.

https://cognisant.io/privacy covers personal data; https://cognisant.io/ai-policy covers inputs to generative AI (Art. 2). Outputs are covered only by our rule that a person reviews them (Art. 1). Both pages are in Japanese.

Not applicable.

We publish our AI Use Policy ( https://cognisant.io/ai-policy ), information security policy ( https://cognisant.io/security-policy ) and a page on how we use AI and AI agents in our own work ( https://cognisant.io/dx ). All are in Japanese.

We are a company of five people, including the Representative Member, and have no board. The Representative Member decides on AI risk and is finally accountable, and serves as the person responsible for information security and personal data. There is no separate advisory body. Policies and rules are reviewed every April and after incidents or changes in laws, customer requirements or the services we use. We have not yet assigned an independent reviewer; we will decide this as part of our information security management system preparation.

On joining, or at the start of a contract, staff are told about our internal rules and the three public policies. We started recording acknowledgements in November 2026. Staff can raise concerns directly with the Representative Member. Material focused on the reliability and risks of generative AI is in preparation.

Incidents and events go into a log kept on an access-controlled shared drive, with what happened, the handling and what will change. Staff report to the Representative Member within one hour. Affected customers are told promptly, and breaches of personal data are reported to the regulator within the legal time limits. The procedure does not yet list AI-specific incidents such as wrong output or prompt injection; we are adding them.

We publish articles in Japanese on evaluating and improving AI systems, including how to build evaluation sets and how to find where accuracy is lost ( https://cognisant.io/insights ).

• ISO/IEC standards (e.g., ISO/IEC 27001, ISO/IEC 42001)
• Other international or regional standards (please specify): Japan's AI Guidelines for Business Ver1.2

We used Appendix 7 of Japan's AI Guidelines for Business Ver1.2 for a self-assessment in November 2026 and will repeat it every year. We are preparing an information security management system based on ISO/IEC 27001, and we plan to consider ISO/IEC 42001 after that. These are self-assessments and preparation, not certifications.

On customer systems, how users are told is agreed with the customer for each project. A company-wide rule for on-screen AI labels is not yet in place; we are adding it to our project-start checklist. Our own public tools (an AI diagnosis tool and a chat assistant) are being reviewed for on-screen AI labels.

This is practice-based work rather than formal research. We develop evaluation methods in our own work, such as measuring each processing stage separately and testing with paraphrased questions. Parts of this are shared in our public articles. We have no dedicated research team.

We publish plain-language articles in Japanese on what AI systems can and cannot do in business use ( https://cognisant.io/insights ). We do not run programs aimed at underserved communities.