Resources
Blog
Field notes on governing AI agents at runtime: MCP policy, spend controls, approvals, audit trails, and real incidents.
Featured
5 postsA Protocol Revision Can Switch Off Your Guardrails Without Raising a Single Error
MCP 2026-07-28 made the protocol stateless. We audited Helio against it and filed sixteen issues. The ones that mattered were not the ones that broke loudly, they were the controls that kept reporting success while enforcing nothing.
Oli Guei A Proxy Cannot Guard Its Own Config File: Where Helio's Enforcement Boundary Actually Is
We found that a process sharing the proxy's user could rewrite Helio's policy live, lift its dashboard secret, and reload a tampered file through a read-only mount. Here is what we reproduced, what shipped in v0.13.1 and v0.14.0, and what the default install still does not defend.
Oli Guei Context Is Not a Safety Boundary: Where Agent Guardrails Actually Have to Live
We verified every incident we could find where AI agents or the tools they call caused real harm. The constraints that failed all lived inside the model's context - and a context window is a buffer, not a boundary.
Oli Guei Malicious npm Packages With Valid SLSA Provenance: Inside the TanStack Attack
On 11 May 2026, malicious @tanstack npm packages shipped with valid SLSA provenance (CVE-2026-45321). Why every signature check passed - and what it means for supply-chain trust.
Oli Guei MCP Agent Governance: The Missing Layer Between Your Agent and Disaster
Your MCP agent can call any tool it can reach, and nothing deterministic checks what it does. The AI agent governance layer most teams skip - and how to fill it.
Oli Guei All posts
A Protocol Revision Can Switch Off Your Guardrails Without Raising a Single Error
MCP 2026-07-28 made the protocol stateless. We audited Helio against it and filed sixteen issues. The ones that mattered were not the ones that broke loudly, they were the controls that kept reporting success while enforcing nothing.
A Proxy Cannot Guard Its Own Config File: Where Helio's Enforcement Boundary Actually Is
We found that a process sharing the proxy's user could rewrite Helio's policy live, lift its dashboard secret, and reload a tampered file through a read-only mount. Here is what we reproduced, what shipped in v0.13.1 and v0.14.0, and what the default install still does not defend.
Context Is Not a Safety Boundary: Where Agent Guardrails Actually Have to Live
We verified every incident we could find where AI agents or the tools they call caused real harm. The constraints that failed all lived inside the model's context - and a context window is a buffer, not a boundary.
Malicious npm Packages With Valid SLSA Provenance: Inside the TanStack Attack
On 11 May 2026, malicious @tanstack npm packages shipped with valid SLSA provenance (CVE-2026-45321). Why every signature check passed - and what it means for supply-chain trust.
MCP Agent Governance: The Missing Layer Between Your Agent and Disaster
Your MCP agent can call any tool it can reach, and nothing deterministic checks what it does. The AI agent governance layer most teams skip - and how to fill it.
No posts match that filter.
Put the gate in. Then let them run.
Install the open-source proxy, watch real traffic, then add the rules your agents need.