Engineering Journal
Engineering Insights
Technical engineering articles on AI trust infrastructure, authorization architecture, evidence systems, and operational security patterns. Reviewed technical analysis from the Vallum engineering team.
Each article includes a maturity indicator showing whether the described capability is implemented and tested, supported by integration, planned, or conceptual.
ArchitectureImplemented & TestedJune 28, 2026·14 min read
Scoped Authorization for AI Agent Operations
Why ambient authority fails for autonomous systems and how external authorization boundaries address the gap
AI agents operating with ambient authority (permissions defined within their prompt context or inherited from a service account) create a fundamental security challenge: any instruction manipulation can potentially access the full permission set. This article examines the architectural pattern of external scoped authorization — where permissions are enforced outside the agent's instruction context, bound to specific tenants and operations, and receipted for audit purposes. We describe how Titan Handshake implements this pattern, the tradeoffs involved (latency, integration complexity, coverage gaps for non-integrated operations), and the failure modes that remain even with external authorization in place.
V
Vallum Engineering
Architecture
authorizationscoped-authoritytitan-handshakefail-closed
SecurityImplemented & TestedJune 14, 2026·16 min read
Tenant-Scoped Evidence Boundaries in Multi-Tenant AI Infrastructure
Architectural approaches to preventing cross-tenant evidence exposure without relying solely on application-layer filtering
Multi-tenant AI systems that store evidence, receipts, or audit records face a specific isolation challenge: a bug in tenant filtering could expose one tenant's operational evidence to another. Application-layer filtering alone creates an unbounded attack surface because every query site must correctly enforce isolation. This article describes the tenant-scoped evidence boundary pattern implemented in Verified Ops, where tenant isolation is enforced at the authorization layer (Titan Handshake) rather than relying on application-level query filters. We discuss what this approach protects against, what it does not protect against (data in systems outside Vallum), and the integration requirements for achieving meaningful isolation.
V
Vallum Engineering
Security
tenant-isolationevidence-boundariesverified-ops
SecurityImplemented & TestedMay 30, 2026·12 min read
External Authorization as a Containment Strategy for Prompt Injection
Limiting the blast radius of instruction manipulation by separating permissions from context
Prompt injection remains a significant vulnerability in deployed AI agents. Most defenses focus on preventing injection (input filtering, instruction hardening) or detecting it (output scanning). This article examines a complementary approach: accepting that injection may succeed at the instruction level, but limiting what a compromised agent can actually do. When permissions are enforced externally via Titan Handshake (not within the agent's context), a successful injection cannot escalate privileges for handshake-protected operations. We discuss the coverage model (only operations routed through Handshake are protected), the failure modes that remain (unprotected operations, social engineering of human approvers), and why this is a containment strategy rather than a prevention guarantee.
V
Vallum Engineering
Security
prompt-injectioncontainmentauthorizationblast-radius
ComplianceImplemented & TestedMay 15, 2026·16 min read
Receipted Execution: Producing Audit Evidence for AI Operations
Generating structured decision records before execution to support compliance documentation requirements
Enterprise AI systems face a growing audit gap: regulators and auditors require documentation of automated decisions, but many AI operations execute without producing structured evidence of what was decided, why, and under what authority. This article describes the receipted execution pattern, where Titan Handshake produces an immutable authorization receipt before a protected operation executes. The receipt captures the decision context (tenant, operation, policies evaluated, outcome) and becomes part of the evidence chain assembled by Titan Verify. We discuss how this pattern supports documentation requirements in SOC2 CC7.2, EU AI Act Article 14, and similar frameworks — while being clear that receipted execution alone does not guarantee compliance (which requires comprehensive organizational controls).
V
Vallum Engineering
Compliance Engineering
auditreceiptsevidence-packagescompliance
Threat AnalysisConceptualApril 28, 2026·10 min read
Metadata Leakage in AI Agent Tool Calls: Risk Assessment
How AI agents may inadvertently disclose operational context through outbound API communications
AI agents making outbound API calls may include operational metadata (user context, business state, internal identifiers) in request parameters, headers, or payloads. Individually innocuous, aggregated metadata from agent tool calls can reveal customer activity patterns and business signals to third-party services. This article assesses the risk surface, categorizes the types of metadata commonly leaked, and discusses potential mitigation strategies. We note that Vallum's current capabilities (receipting outbound operation authorization) provide audit evidence of what was approved, but do not inspect or filter outbound request content — which is an application-layer or network-layer responsibility. We discuss this limitation honestly and identify where additional controls are needed.
V
Vallum Engineering
Threat Analysis
metadatatool-callsdata-leakagerisk-assessment
Stay informed
New technical articles are published periodically. Each article undergoes internal engineering review before publication. For early access to analysis and architecture previews, request documentation access.
Request Documentation Access