Blog
Notes on our work and on software for public institutions.
Open-Source AI Needs a Deliberate Governance Model
Open models offer control and transparency while shifting more operational responsibility to adopters.
Generative AI Makes Data Governance an Engineering Requirement
AI quality and safety depend on clear ownership, lineage, access and retention for source data.
Vector Databases Need Conventional Security Controls
Embeddings and vector stores can expose sensitive meaning even when original documents are hidden.
Synthetic Data Is Useful, but It Is Not Automatically Private
Synthetic datasets need utility, leakage and representativeness testing before use.
Privacy-Enhancing Technologies for Collaborative AI
Federated learning, differential privacy and secure computation can reduce raw-data sharing.
Confidential Computing Protects AI Data While in Use
Hardware-backed trusted execution environments can reduce exposure during sensitive AI processing.
Edge AI Can Improve Privacy—If Data Flows Stay Local
On-device inference reduces cloud exposure but introduces model, device and update risks.
WebAssembly Is Expanding at the Cloud Edge
WebAssembly offers portable, fast-starting workloads with a constrained execution model.
Platform Engineering Uses Golden Paths to Improve Delivery
Internal platforms can combine developer speed with consistent security and operational controls.
DevSecOps Must Adapt to AI-Accelerated Development
Faster code generation requires automated evidence, policy enforcement and risk-based human review.
Memory-Safe Languages Reduce Entire Vulnerability Classes
Adopting memory-safe languages can prevent common corruption flaws in new and critical code.
Secrets Do Not Belong in CI/CD Variables Forever
Modern delivery pipelines should use short-lived identity federation instead of static cloud keys.
Cloud Credential Theft Requires Identity-Centered Detection
Defend cloud environments by reducing standing privilege and detecting abnormal token use.
Ransomware Resilience Depends on Tested Recovery
Immutable backups, identity protection and rehearsed restoration reduce ransomware impact.
API Security Failures Often Begin with Authorization
Strong API security requires object-level authorization, schema validation and abuse monitoring.
Signed Containers and SLSA Improve Software Supply-Chain Trust
Build provenance and artifact signatures help teams verify what was built, where and from which source.
A Practical Kubernetes Security Baseline
Reduce Kubernetes risk with workload identity, policy, image controls and constrained runtime privileges.
Zero Trust for Cloud-Native Systems Is More Than Network Segmentation
Cloud-native Zero Trust combines workload identity, policy, telemetry and least privilege.
Secure by Design Changes Who Owns Cybersecurity Risk
Secure-by-design products make safe defaults and vulnerability reduction core business requirements.
An SBOM Is Useful Only When Connected to Operations
Software bills of materials should drive vulnerability response, procurement and incident investigation.
Passkeys Move Authentication Beyond Reusable Passwords
Passkeys reduce phishing and credential-reuse risk through public-key authentication.
Post-Quantum Cryptography Migration Should Start with Discovery
Organizations should inventory cryptography and prepare for standards-based post-quantum transitions.
FinOps for Generative AI: Control Cost Without Blocking Innovation
Token, tool and infrastructure costs need workload-level measurement and budgets.
AI Observability Starts with Complete Agent Traces
Agent traces connect prompts, retrieval, decisions, tool calls and outcomes for debugging and governance.
Computer-Use Agents Need Transaction-Level Guardrails
Browser and desktop agents should operate with constrained sessions, visible actions and approval gates.
Multimodal AI Expands Both Capability and Attack Surface
Images, audio and video introduce new injection, privacy and content-validation risks.
Why Small Language Models Matter for Enterprise AI
Smaller models can improve cost, latency, privacy and operational control for focused workloads.
Production LLM Evaluation Must Go Beyond Accuracy
Evaluate groundedness, safety, latency, cost and task completion using representative data.
How to Build an AI Red-Team Program for Production Systems
AI red teaming should test the complete application, including tools, data, users and operational controls.
Shadow AI Is an Identity and Data-Governance Problem
Unapproved AI tools can expose business data and create unmanaged identities, connectors and retention paths.
AI Model Supply Chains Need Provenance and Integrity Checks
Protect model weights, adapters, datasets and dependencies from tampering and substitution.
Securing Model Context Protocol Integrations
MCP integrations should use explicit trust boundaries, scoped credentials and tool-level authorization.
AI Agent Memory Is a Security Boundary
Persistent agent memory needs provenance, retention limits and user controls.
Secure RAG Pipelines: Five Controls That Matter
Retrieval-augmented generation needs document trust, access control and grounded-output evaluation.
Preventing Sensitive Data Leakage in LLM Applications
A layered approach to protecting secrets, personal data and proprietary information in AI systems.
Prompt Injection Defense Requires More Than Better Prompts
Effective prompt-injection defense combines content isolation, permissions, validation and monitoring.
How to Secure AI-Generated Code Before It Reaches Production
AI coding speed must be matched by dependency checks, tests, review and runtime controls.
Agentic Source-Code Review Needs Skeptical Validation
Structured agentic review can accelerate vulnerability discovery when paired with expert validation.
Multi-Agent Systems Are Changing Vulnerability Discovery
Microsoft research demonstrates how specialized AI agents can collaborate to find and validate software flaws.
The Security Operations Center Is Becoming Agentic
Agentic security operations combine shared telemetry, reasoning and controlled action to reduce response time.
Adversarial AI Is Moving from Prompts to Agentic Attacks
Google threat research shows attackers adopting automated agent workflows and targeting AI assets.
Code Scanning Is Expanding to Indirect Prompt Injection
GitHub code scanning now detects AI-application risks such as indirect prompt injection and excessive agency.
Zero Trust for AI Agents: A Practical Starting Point
Apply verify explicitly, least privilege and assume breach to AI agents, memory, tools and data.
Why AI Security Must Be Continuously Monitored
New NIST research supports continuous testing and updating instead of one-time AI guardrails.
NIST's 2026 AI Agent Security Findings: What Teams Should Do
NIST's review of AI-agent security highlights novel threats and the need to adapt established cybersecurity practices.
Managed AI Agents Need Strong Sandbox Boundaries
Managed agent platforms simplify execution, but secure sandboxing and permission design remain essential.
Gemma 4 and the Rise of Local Multimodal AI
Gemma 4 12B shows how capable multimodal models are moving onto laptops and edge devices.
Gemini 3.7 Flash Brings Faster Intelligence to Coding Agents
Google's Gemini 3.7 Flash targets coding and agent workflows with improved intelligence, speed and pricing.
What GPT-6 Astra Means for Enterprise AI Systems
A practical look at GPT-6 Astra's reasoning, computer-use and professional-work capabilities for enterprise teams.
GPT-6.1 Sol and the Shift to Multi-Agent Workflows
OpenAI's GPT-6.1 Sol brings lower-cost complex reasoning and beta multi-agent delegation to production workflows.
Random sampling of insurance claims for Nepal's Health Insurance Board
How an openIMIS solution built in Nepal during COVID-19 helped process thousands of claims a day, and became part of the core product.