Cursor Agent Deleted a Prod Database in 9 SecondsFEATURED
Cursor agent reportedly wiped a Railway production database and backups in one API call. Prompts aren't permissions — agents need pre-execution gates.
The Cycles Blog
Engineering insights, product updates, and best practices for AI agent runtime authority, exposure, and cost control.
Tools you can run on your own numbers: Cost Calculator (Claude vs GPT) → · Blast Radius Risk Calculator → — every configuration produces a shareable URL. Many of the posts below link to a calculator pre-loaded with the post's specific scenario.
New to Cycles? Read these posts in order to understand runtime authority from the ground up.
- What Is Runtime Authority for AI Agents?
- The True Cost of Uncontrolled AI Agents
- How Much Do AI Agents Actually Cost?
- AI Agent Cost Management: The Complete Guide
- AI Agent Budget Control: Enforce Hard Spend Limits
Ready to try Cycles? Jump to the End-to-End Tutorial.
Agents Are Cross-Cutting. Your Controls Aren't.FEATURED
· Albert Mavashev · 9 min read
Agents span providers, tools, tenants, and workers. Learn why governance needs cross-cutting budget authority plus application-side authorization at runtime.
State of AI Agent Governance 2026FEATURED
· Albert Mavashev · 14 min read
AI agents moved to production faster than governance kept up. This is the state of enforcement, regulation, and incidents in 2026 — and what comes next.
AI Agent Risk Assessment: Score, Classify, EnforceFEATURED
· Albert Mavashev · 17 min read
Classify AI agent actions by blast radius, assign consistent risk scores, and translate each assessment into enforceable runtime budget controls in practice.
The State of AI Agent Incidents (2026)FEATURED
· Albert Mavashev · 14 min read
Documented AI agent incidents and failure patterns — runaway costs, action misfires, security exploits, multi-agent cascades — scored by cost and blast radius.
Safe Lease Heartbeats for AI AgentsNEW
· Albert Mavashev · 9 min read
Why safe agent lease heartbeats require server-authoritative remaining TTL, monotonic timing, bounded retry windows, and fresh idempotent replay responses.
Proving a Critical-Path Agent Safety GateNEW
· Albert Mavashev · 8 min read
How Cycles tested a critical-path agent gate for latency, saturation, fail-closed behavior, ledger correctness, and durable recovery across four SDKs.
MCP 2026-07-28: What Agent Operators Must ChangeNEW
· Albert Mavashev · 7 min read
The MCP 2026-07-28 release candidate removes sessions and adds routing headers. What operators should prepare before the final specification is published.
Prove You Had Controls: Agent Liability in 2026
· Albert Mavashev · 5 min read
The CMA says you're liable for what your agents do. Contracts exclude the damages. Insurers ask for oversight evidence. Signed decision records serve all three.
Announcing Cycles Budget Guard for Claude Code
· Albert Mavashev · 5 min read
Cycles Budget Guard for Claude Code adds dispatch-path enforcement, blocks denied gated tools before they run, and safely retries recorded-call settlement.
OWASP Agentic Top 10: What Can Be Enforced?
· Albert Mavashev · 6 min read
OWASP's June 2026 data says prompt injection touches six of ten agentic risks. Which ASI categories a pre-execution authority layer can actually enforce.
Lethal Trifecta, Rule of Two, and Metered Authority
· Albert Mavashev · 5 min read
Willison's lethal trifecta and Meta's Rule of Two both subtract agent capabilities. A third option: keep all three legs and meter the most dangerous one.
Agent Identity Is Not Agent Authority
· Albert Mavashev · 6 min read
Okta, Microsoft Entra, and A2A now give AI agents identities. Identity answers who an agent is — not what it may do next, how much, or at whose expense.
The Moltbook Exposure: 1.5M Agent Tokens, No Gates
· Albert Mavashev · 7 min read
Wiz found 1.5M agent API tokens exposed on Moltbook via a hardcoded Supabase key. What it teaches about agent identity, credentials, and runtime authority.
The First Budget to Put on a Production Agent
· Albert Mavashev · 6 min read
The first production agent budget should be small, scoped, and tied to one painful failure mode. Choose tenant, run, or tool budgets based on blast radius.