Powerful enough to find real bugs. Controlled enough to trust.
Letting an AI attack your systems should make you ask hard questions. Here are our answers: what we test, how we limit it, how we use AI, and who is accountable.
Network-enforced scope
Each engagement runs on a dedicated scanner whose outbound traffic is restricted to the hosts in your signed scope. Out-of-scope requests are blocked and logged.
Safe testing tiers
Default checks are non-destructive. Payloads that could modify data, exhaust resources or trigger outages are disabled unless you explicitly opt in, in writing.
Rate limits & testing windows
Per-host request limits and optional time windows keep load predictable. We never run denial-of-service tests.
Kill switch
You can halt all testing instantly from the portal or by calling us. Agents stop mid-run.
Static, published egress IPs
Allow-list them on your WAF. We do not attempt to bypass or evade your defensive controls.
Secrets never reach the model
Test credentials are stored in a vault and used by an auth broker. AI agents receive short-lived session tokens only.
Complete audit trail
Every request, tool call and agent decision is logged with hash-chained integrity, available to you on request.
Human approval gates
New assets discovered during recon are reported, never tested, until you approve adding them to scope.
Data retention & deletion
Engagement data is retained only as long as your contract requires, then destroyed, with a deletion certificate on request.
How we test
Our testing follows established, publicly documented methodologies so that auditors and your security team can understand and verify our approach:
- OWASP Web Security Testing Guide (WSTG) for web application test cases
- OWASP Top 10:2025 and OWASP API Security Top 10 (2023) for risk categorisation
- OWASP ASVS 5.0 as a verification reference
- PTES and NIST SP 800-115 for engagement phases, from pre-engagement to reporting
- CVSS 3.1 and 4.0 for severity, computed from the vector, never estimated by an AI
- CWE identifiers for every finding, validated against MITRE’s list
A finding is reported only if it is reproducible and supported by captured evidence. Every Critical and High is reproduced by a human before it is delivered.
AI disclosure
We believe you have a right to know exactly how AI is involved in testing your systems.
- What AI does: plans and executes test steps, analyses responses, attempts exploitation within scope, de-duplicates results and drafts report text.
- What humans do: agree scope and rules of engagement, verify serious findings, confirm severity, review every report, sign attestation letters, and perform expert-led testing in every Audit-Ready Pentest.
- Models: we use open-weight large language models, either self-hosted or accessed through providers under zero-data-retention terms. A self-hosted-only option is available for confidential environments.
- Training: your data is never used to train or fine-tune models.
- Sub-processors: every infrastructure and model provider that may process engagement data is named in your authorisation letter and data processing agreement before testing begins.
Authorised testing only
We only test assets that the customer owns or is authorised to test, under a signed authorisation letter and rules of engagement. We do not test third-party services (for example your SaaS vendors or cloud provider control planes) without that provider’s permission, and we follow the published penetration testing policies of major cloud providers.
Reporting a vulnerability in TensorShield
If you believe you have found a security issue in our website or services, please see our responsible disclosure policy or email tensorshield@gmail.com.
Have a question your CISO will ask?
We are happy to walk your security team through our guardrails, data handling and methodology before you sign anything.