Access reviews approve the breach
Blog · November 17, 2026 · Karim Hijazi
An access review can tell you an agent is allowed to reach HubSpot. It does not tell you whether the MCP server that agent installed is reading ~/.npmrc.
That is the gap. IT tools now run access reviews, provisioning and audits for AI agents the same way they do for people. They treat those agents as non-human identities. They answer a real question: who can reach what. They do not answer the question that matters once a tool is already allowed. Does that tool actually do what it claims?
I keep hearing the first question treated as if it were both. It is not.
1. The review already said yes
Access governance is built to approve trust. A person or an agent is in the directory. An app is on the approved list. The owner signs the review. The ticket closes. The file for the audit is complete.
That process is doing its job when it says yes to a known vendor, a known app or a known agent. Breaches arrive through exactly that kind of yes.
On November 13, 2026, the FBI published FLASH-20261113-001. It describes ARTEX, an open-source, LLM-driven intrusion framework used in the October ConnectWise Cloud Backup supply-chain compromise. The intrusion reached more than 2,250 downstream Microsoft 365 tenants. In some cases it went from targeting to privileged access in under three hours.
Look at that through an access review. The software was already in the environment. The identities were already approved. The path to those tenants was already trusted. A review of who could reach what would have had every reason to pass.
This week made the same point in another form. A compromised Optimizely script served attacker code on about 178 organizations' sites, including CNN, BBC and NPR. Every page trusted one outside script. The script was there for a reason. The trust was already granted. That was the opening.
The reviews are not designed to run the component and watch what it does. They are designed to confirm that a trusted thing is allowed. When the trusted thing is the path in, the review has already done the only thing it can do. It approved it.
2. Why the access layer gets funded first
If you sit in IT, this is not a hard call. Access governance already has a budget line. Audits already require the reviews. Adding agents as non-human identities extends a process you already run. It is easy to buy, and easy to explain to whoever owns the audit.
The tool-behavior question is a newer ask, and it is not the one the audit already requires. So the access layer wins the first conversation.
I understand why. I also think it is a mistake to treat that purchase as the end of the work. Access governance tells you the agent is allowed to reach the app. It does not tell you what the MCP server or skill behind that access is doing once it is installed. Those are different facts. Only one of them is in the review today.
When a buyer has to choose, they fund the layer the audit already asks for. When a breach shows up, it often comes through a component that same layer would have approved. That is the mismatch.
3. Check the tool, not just the identity
This is the layer we built KillerSkills to cover.
We test MCP servers and agent skills by running them. We call that detonation. The output is a verdict, tied to a published version and its artifact hash. A later release, or a change in the bytes, does not inherit the old verdict.
You can start on your own machine. Run npx vet-skills inventory. It is free and does not need an account. It lists the agents, MCP servers and skills on that machine, and it shows credentials sitting in configs.
The guard watches what happens after that. If an update reaches for a new host, a new credential or a new tool, the guard can hold it. It defaults to observe on first run, and to warn on a change.
Then vet-skills prove shows where credentials were found. It covers six CLI agents: Claude Code, Codex CLI, Gemini CLI, Qwen Code, Grok CLI and Copilot CLI. It names the file and the kind of credential. It does not print the value. It is not proof that the credential is off the machine.
None of this replaces an access review. An inventory will not tell you who should have HubSpot. A verdict will not provision an account. What it will do is answer the question the review never asks: for this version, with this hash, what did the tool actually do?
4. Put the verdict in the review you already run
I do not want companies to throw out the access layer and start over. Keep the reviews. Keep the provisioning. Keep the agents in the same process as everyone else.
Add the missing input. When you review an agent's access to an app, look at the verdict on the tools that agent runs, for the exact version and hash on the machine. If a tool starts reaching for a new host or a credential it did not use before, that fact belongs in the same review as the access grant.
Access governance tells you the agent is allowed in. The behavior layer tells you what the allowed tool is doing. Both are needed. The second should feed the first.
If you want to see what is already on a machine, start here:
npx vet-skills inventory
Free. No account.