RaksHex Research & Analysis
The long-form layer behind RaksHex Incident Intelligence.
The incident database is the evidence layer. This blog is where we unpack the architecture, economics and security lessons behind those records without turning every event into a product pitch.
A forgotten Google Cloud API key was abused for 60,000+ requests and an $18,000+ bill.
A reusable credential, an exposed execution path and a budget that did not act as authorization combined into a costly failure. We separate what RaksHex could reduce from what remains outside its path.
Read analysis →Claude Code users approve 93% of permission prompts. That is a control-design problem.
Anthropic's auto-mode write-up shows why asking humans to approve everything eventually weakens the approval itself. Contextual action policy is a different control boundary.
Read analysis →Cloudflare added hard AI spend limits. Cost control is becoming infrastructure.
Cloudflare AI Gateway can now block requests when a dollar budget is exhausted. We break down what that means for runaway agent spend and why cost enforcement alone is not enough.
Read analysis →Google Cloud is treating AI agents as first-class identities.
Google's latest MCP guidance recommends dedicated agent identities and least privilege in production. That makes identity part of the agent-action security boundary.
Read analysis →Microsoft says MCP needs a policy point before tool execution.
Microsoft's Agent Governance Toolkit targets a specific gap in MCP: policy is not built into the protocol's execution surface. The authorization decision has to live somewhere else.
Read analysis →When an AI coding agent deleted a production database during a code freeze.
The Replit/SaaStr incident remains a clean example of the difference between telling an agent not to act and technically preventing the action.
Read analysis →