| Governance |
|---|
| Three-layer security architecture | Every security question has one of three answers: a platform invariant, your policy, or assurance. Nothing is an exception, and pilot choices are recorded as that customer's policy. | SOC 2CC5.2 Technology controls chosen and built ISO 27001A.5.1 Information security policies NIST CSFGV.PO Govern: policy GDPRArt. 25 Data protection by design and by default ISO 42001A.2 Policies for AI |
|---|
| Threat model with named defences | Nineteen named threats, each with the defence that stops it and the signal that detects it, from prompt injection to a stolen phone to a supplier breach. Reviewed when the system changes. | SOC 2CC3.2 Risk identification and analysis ISO 270016.1.2 Information security risk assessment (clause 6.1.2) NIST CSFID.RA Identify: risk assessment NIST AI RMFMAP 5.1 Likelihood and size of impacts identified ISO 42001A.5 Assessing impacts of AI systems |
|---|
| Every requirement traced to a check | All 247 requirements map to a named check of five kinds: test, safety test, evaluation, runtime guard or review. | SOC 2CC4.1 Ongoing evaluation that controls work ISO 27001A.8.29 Security testing before release NIST CSFID.IM Identify: improvement NIST AI RMFMEASURE 2.5 Validity and reliability shown ISO 42001A.6 AI system life cycle |
|---|
| Security review before go-live | A review of sign-in, roles, the tenant wall, secrets, public addresses, event receivers, content security policy, process privileges and red-team results. Every finding is fixed or accepted in writing, with its date. | SOC 2CC4.1 Ongoing evaluation that controls workCC8.1 Change management ISO 27001A.8.29 Security testing before release GDPRArt. 32(1)(d) Regular testing of security measures NIST CSFID.IM Identify: improvement |
|---|
| Excluded high-risk uses | Keepara is not offered for credit, insurance pricing, employment, essential-service eligibility, or any decision with legal effect on a person. Contracts prohibit these uses. | EU AI ActArt. 5 Prohibited AI practicesArt. 6, Annex III High-risk classification GDPRArt. 22 No solely automated decisions with legal effect CCPAADMT regulations California rules on automated decisionmaking technology NIST AI RMFMAP 1.1 Intended purpose and context documented ISO 42001A.9 Responsible use of AI systems |
|---|
| AI literacy and role training | The team building Keepara and the people approving drafts are trained on what the system does, what it cannot do, and how to stop it. | EU AI ActArt. 4 AI literacy SOC 2CC1.4 Competent, trained people ISO 27001A.6.3 Security awareness and training NIST CSFPR.AT Protect: awareness and training NIST AI RMFGOVERN 2.2 AI risk training for staff and partners ISO 42001A.4 Resources, including people and skills |
|---|
| Legal and regulatory register | This control map, kept in the repository beside the architecture, records which requirement each control serves and is regenerated whenever a control changes. | ISO 27001A.5.31 Legal and contractual requirements NIST AI RMFGOVERN 1.1 Legal and regulatory requirements understood GDPRArt. 5(2) Accountability ISO 42001A.2 Policies for AI |
|---|
| Isolation and ownership |
|---|
| Database-enforced tenant wall | Row-level security on every table. A transaction with no business attached raises an error. The isolation suite runs every pair of 50 simulated businesses. | SOC 2CC6.1 Logical access securityC1.1 Protecting confidential information ISO 27001A.8.3 Information access restriction NIST CSFPR.DS Protect: data security GDPRArt. 32(1)(b) Ongoing confidentiality, integrity, availability, resilience CCPA§1798.140(ag) Service provider duties, including no combining across businesses OWASP LLMLLM08 Vector and embedding weaknesses |
|---|
| Tenant-local learning | Lessons come only from your own team's explicit feedback, carry their author, can be reversed, and never leave your business. No model route that trains on inputs can be used. | GDPRArt. 5(1)(b) Purpose limitationArt. 29 Processing only on instructions CCPA§1798.140(ag) Service provider duties, including no combining across businesses ISO 42001A.7 Data for AI systems NIST AI RMFMEASURE 2.10 Privacy risk examined OWASP LLMLLM04 Data and model poisoning |
|---|
| One customer per model call | Each model call carries one customer's content, or one joint meeting's, and the gateway refuses anything else. A fact said by one customer about another is never used for either unless marked shared. | OWASP LLMLLM02 Sensitive information disclosure SOC 2C1.1 Protecting confidential information ISO 27001A.8.12 Data leakage prevention GDPRArt. 32 Security of processing |
|---|
| You own your data and what is derived from it | Your customers' records, the agent's memory of them and every insight derived from them belong to your business. Keepara processes them only to serve you. | GDPRArt. 28 Processor obligationsArt. 29 Processing only on instructions CCPA§1798.140(ag) Service provider duties, including no combining across businesses ISO 42001A.10 Third-party and customer relationships |
|---|
| Identity and access |
|---|
| Sign-in with two-step verification | Only accounts your admin adds can sign in. Two-step verification is checked and recorded for every user. Repeated refused sign-ins raise an alert. | SOC 2CC6.1 Logical access securityCC6.2 Adding and removing users ISO 27001A.8.5 Secure authenticationA.5.17 Authentication information NIST CSFPR.AA Protect: identity, authentication, access |
|---|
| Least privilege, by role | Every role starts with the least it needs. Widening is an admin action, logged where nobody can edit it. | SOC 2CC6.3 Role-based access and least privilege ISO 27001A.5.15 Access controlA.5.18 Access rightsA.8.2 Privileged access rights NIST CSFPR.AA Protect: identity, authentication, access GDPRArt. 25 Data protection by design and by default |
|---|
| Operators see IDs only | People running the service work from IDs and counts. Customer content is outside the operator role, and the system-admin role sees no business's content. | SOC 2CC6.3 Role-based access and least privilegeC1.1 Protecting confidential information ISO 27001A.8.2 Privileged access rightsA.8.11 Data masking GDPRArt. 32(4) Staff act only on instructions |
|---|
| Sessions ended in one step | Session limits on every sign-in. Removing a user ends their sessions at once. An admin ends all of one user's sessions in a single action, for a lost phone or a suspected stolen login. | SOC 2CC6.2 Adding and removing users ISO 27001A.5.18 Access rightsA.8.5 Secure authentication NIST CSFPR.AA Protect: identity, authentication, access |
|---|
| Process privilege separation | Each running process holds only its own keys and its own database role. The learning worker cannot write facts or check settings. The backup job cannot read the backup private key. | SOC 2CC6.3 Role-based access and least privilege ISO 27001A.5.3 Segregation of dutiesA.8.2 Privileged access rights NIST CSFPR.AA Protect: identity, authentication, access OWASP LLMLLM06 Excessive agency |
|---|
| Access reviews | The list of users, roles, named powers and system admins is reviewed at each release and before go-live. Changes go to logs nobody can edit. | SOC 2CC6.2 Adding and removing users ISO 27001A.5.18 Access rights NIST CSFGV.RR Govern: roles and responsibilities |
|---|
| Data protection |
|---|
| Encryption in transit and at rest | Connections only over TLS. Databases and backups encrypted at rest. Backups are encrypted on the server before they leave, with the private key held outside every provider. | SOC 2CC6.1 Logical access securityCC6.7 Protecting data in transmission ISO 27001A.8.24 Use of cryptography NIST CSFPR.DS Protect: data security GDPRArt. 32(1)(a) Encryption and pseudonymisation CCPA§1798.100(e) Reasonable security§1798.150 Liability for breaches from unreasonable security |
|---|
| Secrets management | Secrets live in encrypted environment settings, one set per process, never in code or its history. A secret scan runs on every change. Any exposed secret is rotated at once. | SOC 2CC6.1 Logical access security ISO 27001A.5.17 Authentication informationA.8.24 Use of cryptography NIST CSFPR.DS Protect: data security OWASP LLMLLM07 System prompt leakage |
|---|
| Least data in | A keep-list filter runs before storage. Anything not about a kept customer is dropped unread and counted. Card numbers and unrelated records never enter. | GDPRArt. 5(1)(c) Data minimisationArt. 25 Data protection by design and by default CCPA§1798.100(c) Collection reasonably necessary and proportionate SOC 2P3.1 Collection limited to the purpose ISO 27001A.5.34 Privacy and protection of personal data |
|---|
| Read-only connections, narrowest scopes | Every connection to your tools is read-only, using the narrowest permission each provider offers. A stolen token could read, never change or send. | SOC 2CC6.1 Logical access securityCC6.3 Role-based access and least privilege ISO 27001A.5.15 Access control NIST CSFPR.AA Protect: identity, authentication, access OWASP LLMLLM06 Excessive agency |
|---|
| No content in logs, alerts or monitoring | System logs, alerts, error reports and notifications carry IDs and counts only, enforced by a field allow-list and tested with planted values. | ISO 27001A.8.11 Data maskingA.8.12 Data leakage preventionA.8.15 Logging SOC 2C1.1 Protecting confidential information GDPRArt. 25 Data protection by design and by default OWASP LLMLLM02 Sensitive information disclosure |
|---|
| Residency in your chosen region | Your data lives in the region you choose, including several locations with their own residency terms. Transfers outside it rely on recognised safeguards: EU Standard Contractual Clauses, the EU-US Data Privacy Framework where the provider is certified, and the UK transfer addendum. | GDPRArt. 44-46 International transfers with safeguards ISO 27001A.5.23 Security of cloud services NIST CSFGV.SC Govern: supply chain risk |
|---|
| Retention you set | Import depth and retention windows are your settings. Facts weigh less as they age, and an old fact is never presented as current. | GDPRArt. 5(1)(e) Storage limitation SOC 2P4.2 Retention of personal information ISO 27001A.8.10 Information deletion |
|---|
| Sensitive matters handled with care | Sensitive facts are labelled and restricted. They reach a customer only in that customer's own words. A mixed session's raw text is stored restricted and shown on no screen. | GDPRArt. 9 Special categories of personal data CCPA§1798.121 Right to limit sensitive personal information SOC 2C1.1 Protecting confidential information |
|---|
| AI safety |
|---|
| Untrusted content boundary | Source text enters a model only inside a labelled untrusted region. Extraction returns a fixed schema. The reasoning step has no tools and no network. Drafting never sees raw source text. | OWASP LLMLLM01 Prompt injectionLLM05 Improper output handling NIST AI RMFMEASURE 2.7 Security and resilience evaluated EU AI ActArt. 15 Accuracy, robustness, cybersecurity (high-risk duty, followed by choice) ISO 27001A.8.26 Application security requirements ISO 42001A.6 AI system life cycle |
|---|
| Seven checks on every draft | Nothing invented; nothing from another customer; nothing stale; no false familiarity; within contact limits; no harm to a sensitive matter; your own rules. In that order, with no shortcut. | OWASP LLMLLM09 MisinformationLLM02 Sensitive information disclosure SOC 2PI1.3 Accurate, complete processing NIST AI RMFMEASURE 2.5 Validity and reliability shown EU AI ActArt. 15 Accuracy, robustness, cybersecurity (high-risk duty, followed by choice) ISO 42001A.6 AI system life cycle |
|---|
| A person sends by default | Every action type starts at Propose. Autonomy is granted by the business one action type at a time, after a track record, with a recall window. Nothing widens its own permissions. | GDPRArt. 22 No solely automated decisions with legal effect EU AI ActArt. 14 Human oversight (high-risk duty, followed by choice) NIST AI RMFMAP 3.5 Human oversight defined ISO 42001A.9 Responsible use of AI systems OWASP LLMLLM06 Excessive agency |
|---|
| AI-assistance disclosure on by default | Messages drafted with Keepara carry a line saying so, by default, and the business may change it where its own law allows. | EU AI ActArt. 50 Transparency for AI-generated content NIST AI RMFMEASURE 2.8 Transparency and accountability examined ISO 42001A.8 Information for interested parties |
|---|
| Approved model routes only | A model route can be called only once its provider, region, retention and training terms are recorded. Routes that keep nothing are preferred. An answer from an unlisted provider raises an alert. | GDPRArt. 28(2) Prior authorisation of sub-processorsArt. 28(4) Same duties flow down to sub-processors SOC 2CC9.2 Vendor and partner risk ISO 27001A.5.19 Security in supplier relationships NIST AI RMFGOVERN 6.1 Third-party AI risk policies OWASP LLMLLM03 Supply chain ISO 42001A.10 Third-party and customer relationships |
|---|
| Spend bounded per decision | Each wake has a fixed budget and ends with what it has. Jobs are safe to run twice and make one draft per reason. Spending is watched hourly, with an alert at five times the expected rate. | OWASP LLMLLM10 Unbounded consumption SOC 2A1.1 Capacity management ISO 27001A.8.6 Capacity management |
|---|
| Identity matching with a person in the loop | Uncertain matches wait for a person. Nothing is drafted for an unconfirmed customer. Every merge can be split. | GDPRArt. 5(1)(d) Accuracy SOC 2PI1.3 Accurate, complete processing NIST AI RMFMEASURE 2.5 Validity and reliability shown OWASP LLMLLM09 Misinformation |
|---|
| Evaluations and red-teaming on every change | Leak, fabrication, tone and prompt-injection sets run on every prompt or model change. Leak and injection sets must pass at 100%. | NIST AI RMFMEASURE 2.7 Security and resilience evaluatedMANAGE 4.1 Monitoring after deployment EU AI ActArt. 15 Accuracy, robustness, cybersecurity (high-risk duty, followed by choice) SOC 2CC8.1 Change management ISO 27001A.8.29 Security testing before release OWASP LLMLLM01 Prompt injectionLLM02 Sensitive information disclosure ISO 42001A.6 AI system life cycle |
|---|
| Decision trail for every draft | What woke the agent, the facts it used, the rules that fired, each check's result, the model and prompt version, who approved, and what was sent. | EU AI ActArt. 12 Record-keeping (high-risk duty, followed by choice)Art. 13 Information for deployers (high-risk duty, followed by choice) GDPRArt. 5(2) Accountability NIST AI RMFMEASURE 2.8 Transparency and accountability examined SOC 2PI1.3 Accurate, complete processing ISO 42001A.8 Information for interested parties |
|---|
| Logging and monitoring |
|---|
| Every content read logged | Written in the same database transaction as the read. A read whose log line cannot be written fails. Logs cannot be edited. | SOC 2CC7.2 Monitoring for anomalies NIST CSFDE.CM Detect: continuous monitoring GDPRArt. 30(2) Processor records of processingArt. 32 Security of processing |
|---|
| Alerts on the signals that matter | Bursts of refused sign-ins, an unlisted model provider answering, runaway spend, and any leak or invented fact reported by a user. | SOC 2CC7.2 Monitoring for anomaliesCC7.3 Evaluating security events ISO 27001A.8.16 Monitoring activities NIST CSFDE.AE Detect: adverse event analysis NIST AI RMFMANAGE 4.1 Monitoring after deployment |
|---|
| Verified inbound events, listed hosts only | Every pushed event is verified by signature or channel token before anything else. Keepara fetches only from each source's own listed hosts, so an altered link cannot reach an internal service. | SOC 2CC6.6 Protection at the system boundary ISO 27001A.8.20 Network security NIST CSFPR.IR Protect: infrastructure resilience |
|---|
| Content shown as text, never as code | All customer content is rendered as plain text under a strict content security policy, so a note cannot run script in the app. | OWASP LLMLLM05 Improper output handling ISO 27001A.8.26 Application security requirementsA.8.28 Secure coding SOC 2CC6.8 Preventing malicious code |
|---|
| Secure development |
|---|
| Pinned and scanned dependencies | Exact pinned versions, scanned on every change and nightly. A high or critical flaw fails the build. | SOC 2CC7.1 Detecting vulnerabilities and configuration changeCC8.1 Change management ISO 27001A.8.8 Technical vulnerability managementA.8.28 Secure coding NIST CSFPR.PS Protect: platform security OWASP LLMLLM03 Supply chain |
|---|
| Separate environments, fabricated test data | Development, staging and production are separate. Production data stays in production. Tests run on fabricated businesses and customers. | ISO 27001A.8.31 Separate development, test and productionA.8.33 Test information SOC 2CC8.1 Change management GDPRArt. 25 Data protection by design and by default |
|---|
| Reviewed, tested changes | Every change passes the test suite and safety tests in continuous integration before release. Prompt and model changes also pass the evaluation and red-team sets. | SOC 2CC8.1 Change management ISO 27001A.8.25 Secure development life cycleA.8.32 Change management NIST CSFPR.PS Protect: platform security ISO 42001A.6 AI system life cycle |
|---|
| Resilience |
|---|
| Encrypted backups and a practised restore | Nightly encrypted backups, checked by size and checksum. A weekly test restore compares row counts. A practised restore targets a working system within 4 hours, losing at most one day of data. | SOC 2A1.2 Backup and recoveryA1.3 Recovery testingCC7.5 Recovering from incidents ISO 27001A.8.13 Information backupA.5.30 ICT readiness for business continuity NIST CSFRC.RP Recover: recovery plan execution GDPRArt. 32(1)(b) Ongoing confidentiality, integrity, availability, resilienceArt. 32(1)(c) Timely restore after an incident |
|---|
| One-step halt | Any user stops waking, drafting and model calls in one action, within 10 seconds, surviving restarts. Held drafts are checked again after. | NIST AI RMFMANAGE 2.4 Means to stop or disengage the system EU AI ActArt. 14 Human oversight (high-risk duty, followed by choice) SOC 2CC7.4 Responding to incidents ISO 42001A.9 Responsible use of AI systems OWASP LLMLLM06 Excessive agency |
|---|
| Continuity for your data | Your data can be exported in an open format at any time on request, and at exit. If Keepara stops operating, your data is returned and then deleted. | SOC 2CC9.1 Business disruption risk ISO 27001A.5.30 ICT readiness for business continuity GDPRArt. 20 Right to data portabilityArt. 28(3)(g) Delete or return all data at the end |
|---|
| Privacy rights |
|---|
| Exit means gone | All data returned or deleted at exit, including every backup copy, with written confirmation. A deletion register outside the database is re-applied after any restore. | GDPRArt. 17 Right to erasureArt. 28(3)(g) Delete or return all data at the end CCPA§1798.105 Right to delete SOC 2C1.2 Disposing of confidential informationP4.3 Secure disposal of personal information ISO 27001A.8.10 Information deletion |
|---|
| Help with every individual request | Tools to find, export, correct and erase one customer's data, so your business can answer access, correction, deletion and portability requests. | GDPRArt. 15 Right of accessArt. 16 Right to rectificationArt. 17 Right to erasureArt. 20 Right to data portabilityArt. 28(3)(e) Processor helps with data subject requests CCPA§1798.105 Right to delete§1798.106 Right to correct§1798.110 Right to know SOC 2P5.1 Access to personal information |
|---|
| Opt-outs are absolute | A customer who opts out is never proposed for contact again. No setting can override it. Contact caps and spacing limit how often anyone is approached. | GDPRArt. 21(3) Objection to direct marketing is final CCPA§1798.120 Right to opt out of sale and sharing SOC 2P3.1 Collection limited to the purpose |
|---|
| Support for impact assessments | Architecture, data flows, this control map and the threat model are provided to support your data protection impact assessment and risk assessments. | GDPRArt. 35 Data protection impact assessmentArt. 28(3)(h) Information and audits EU AI ActArt. 13 Information for deployers (high-risk duty, followed by choice) ISO 42001A.5 Assessing impacts of AI systems |
|---|
| Suppliers and incidents |
|---|
| Published subprocessor register | Every outside service that touches customer data is named with its role, what it sees and where. Customers are told before a new one is added, and the same duties flow down by contract. | GDPRArt. 28(2) Prior authorisation of sub-processorsArt. 28(4) Same duties flow down to sub-processors CCPA§1798.140(ag) Service provider duties, including no combining across businesses SOC 2CC9.2 Vendor and partner risk ISO 27001A.5.19 Security in supplier relationshipsA.5.20 Security in supplier agreementsA.5.21 ICT supply chain security NIST CSFGV.SC Govern: supply chain risk ISO 42001A.10 Third-party and customer relationships |
|---|
| Incident response | Runbooks for incidents, leaks and account compromise. Contain first, then assess, then tell you without undue delay, with the facts you need for your own legal deadlines. | GDPRArt. 33(2) Processor reports breaches without undue delay CCPA§1798.82 California breach notice: service providers tell the data owner SOC 2CC7.3 Evaluating security eventsCC7.4 Responding to incidents ISO 27001A.5.24 Incident management planningA.5.26 Response to incidentsA.5.27 Learning from incidents NIST CSFRS.MA Respond: incident managementRS.CO Respond: incident communication NIST AI RMFMANAGE 4.3 Incidents and errors communicated |
|---|
| Vulnerability disclosure | A security contact at [email protected], a security.txt file, and good-faith safe harbour for researchers who report a flaw responsibly. | ISO 27001A.6.8 Security event reportingA.8.8 Technical vulnerability management NIST CSFID.RA-08 Identify: receiving vulnerability disclosures SOC 2CC2.3 Communication with customers and other outside parties |
|---|