Effective penetration testing demands authenticated access. Testing modern applications without valid session tokens, test user credentials, or API keys leaves critical business logic, Broken Object Level Authorization (BOLA), and privilege escalation paths completely unexamined. However, handing production-like credentials to an autonomous artificial intelligence engine presents legitimate security concerns for CISOs and compliance teams.
To resolve this fundamental tension, Lorikeet Security engineered Lory with a strict zero-knowledge execution architecture. In this deep dive, we explain the cryptographic mechanisms, runtime memory zeroing, prompt token scrubbing, and strict tenant isolation that ensure customer secrets remain secure throughout autonomous security assessments.
The Secret Management Challenge in Autonomous Pentesting
Autonomous penetration testing platforms face three distinct risks when handling user credentials:
- Model Inference Exposure: If sensitive passwords or session bearer tokens are injected directly into LLM prompts, they risk being logged by LLM model API providers or retained in inference logs.
- Cross-Tenant Contamination: Multi-tenant SaaS architectures must guarantee that secrets provisioned by Company A can never leak into vectors, tools, or memory spaces utilized during Company B audits.
- Post-Execution Persistence: Once an assessment completes, temporary credentials in runtime memory or local disk stores could remain exposed to memory dumps or unauthorized inspection.
AES-256-GCM Envelope Encryption at Rest
Customer credentials configured within the Talon platform or supplied via CI/CD pipelines are never stored in plaintext. Secrets are encrypted using Authenticated Encryption with Associated Data (AEAD) via AES-256-GCM.
Each customer tenant receives unique, dynamically generated Data Encryption Keys (DEKs). Master Key Encryption Keys (KEKs) reside within certified Hardware Security Modules (FIPS 140-2 Level 3 compliant). Key retrieval is strictly logged and audited under SOC 2 Type II Confidentiality controls.
Prompt Token Masking: Scrubbing Secrets Before the LLM
The cardinal rule of Lory architecture is: LLM inference endpoints never see plaintext credentials.
Before Lory generates reasoning prompts or constructs attack plans across Layer 1 or Layer 2 orchestrators, all known credentials undergo deterministic token substitution:
| Secret Class | Actual Customer Secret | Substituted Surrogate Token | Runtime Injection Point |
|---|---|---|---|
| Bearer Auth Header | Bearer eyJhbGciOi...98xQ |
{{LORY_AUTH_TOKEN_ALPHA}} |
Worker cURL / HTTP engine directly |
| Tenant API Key | lsk_live_99f2b8813a0c |
{{LORY_API_KEY_TEST}} |
Isolated CLI sandbox process |
| Test User Password | ComplexPass#2026!Orlando |
{{LORY_USER_PWD_BETA}} |
Headless Chromium authentication form |
| Database Connection | postgres://adm:9f#1@db... |
{{LORY_DB_CONN_SURROGATE}} |
Layer 3 network runner |
The artificial intelligence model reasons solely with the abstract placeholders: "Submit an HTTP POST request to /api/v2/orders including {{LORY_AUTH_TOKEN_ALPHA}} to evaluate if non-admin roles can read billing details." The actual secret is populated exclusively within the isolated worker sandbox runtime just microseconds before the socket transmission occurs.
Lorikeet Security maintains enterprise zero-retention service level agreements (SLAs) with all model providers. Prompts, outputs, and intermediate data are never used for model training, reinforcement learning, or human review by upstream vendors.
Ephemeral Memory Zeroing and Isolated Worker Sandboxes
When Lory dispatches an authenticated test vector (such as an IDOR probe or a JWT privilege escalation check), the execution occurs inside an isolated, short-lived container sandbox:
- Per-Vector Container Isolation: Each vector execution runs in a locked-down container with zero access to sibling vectors or peer tenant networks.
- RAM-Only Decryption: Secrets exist only in transient memory registers, never written to swap partitions or container disk storage.
- Explicit Memory Scrubbing: Upon vector termination, memory buffers containing decrypted tokens are overwritten with zeros (cryptographic memory zeroing) before process exit.
- Instant Sandbox Destruction: Containers are immediately purged upon task completion. No residual state or filesystem layers persist on host hypervisors.
Regional Database Pinning & Compliance Standards
For organizations operating under strict data residency mandates (such as Canadian PIPEDA, European GDPR, or US federal compliance), Lory supports regional tenant pinning:
- US-Only Workspaces: Data, keys, and execution workers are locked strictly to US East (Northern Virginia) and US Central regions.
- Canadian Workspaces: Dedicated encryption envelopes and compute clusters stationed in Montreal and Toronto.
- Audit Alignment: Fully compliant with SOC 2 Type II (Security, Availability, Confidentiality, Privacy), ISO/IEC 27001:2022 Annex A.9 (Access Control) and A.10 (Cryptography), and HIPAA §164.312(a)(2)(iv) encryption mandates.
Enterprise-Grade Security for Autonomous Audits
Experience autonomous penetration testing engineered with the world's most rigorous secret protection and zero-knowledge execution protocols.
