BLOG
Two critical PHP runtime flaws show how AI can accelerate exploit research while patching still depends on maintainers, providers, and operators proving what runs in production.
What Google's universal work agent reveals about persistent memory, delegated identity, MCP tools, model routing, revocation, and enterprise control.
How AI helped investigators trace funds after the 2026 Bitget theft, and where it could improve detection and response without controlling transactions.
What Muse, Apple's Full Disk Access announcement, and agent payments reveal about permission scope, prompt injection, and the evidence behind safety claims.
A minimal reproducible post-mortem on how a helpful coding agent turned destructive mid-loop, and the four deterministic engineering controls that prevent it in CI/CD.
A comprehensive technical breakdown of why prompt-based guardrails fail in production, dissecting five critical agent failure modes and presenting a five-layer deterministic engineering control architecture.
A point-in-time architecture review of CrewAI covering inter-agent trust, permission cascades, loop exhaustion, recovery, and deterministic orchestration controls.
A point-in-time architecture review of OpenHands covering sandbox boundaries, terminal execution risks, prompt injection exposure, and deployment controls.
The DseWiki reports and the OpenAI/Hugging Face incident show why capable AI agents need system-level boundaries: controlled egress, isolated state, trustworthy telemetry, safe exits, and revocable authority.
A foundational guide for engineering teams building AI agents: why front-end safety is not enough, why validators are not safety reviewers, and what secure agent design looks like across design, development, runtime, and governance.
A production incident case study: what Amazon's AI-assisted outages reveal about runaway blast radius, overprivileged agents, validation blind spots, and missing release controls.
