Failure pattern library

The same failures keep recurring

Each incident on record is tagged with the patterns that explain it. A pattern page collects those incidents with their preventive controls. The tagging is our editorial judgment. How it works

Indirect prompt injection

Content the agent read from an untrusted source was treated as instructions.

9 incidents

Excessive permissions

The agent, or a credential it could reach, was able to do more than the task needed.

8 incidents

Missing isolation

Environments, tenants or data that should have been separate were reachable from one another.

7 incidents

Untrusted supply chain

A third-party package, skill, extension or server ran without verification of its source or changes.

7 incidents

Unreviewed change

A change to a prompt, configuration or infrastructure reached production without adequate review, staging or monitoring.

6 incidents

Unsafe command execution

Commands ran without a review or confirmation step that could have stopped them.

4 incidents

Open exfiltration channel

Model output could carry data to an outside party, for example through a link or an image address.

4 incidents

Ungrounded answers

The model stated policies or facts that had no basis in an authoritative source.

3 incidents

Weak guardrails

A user could talk the model out of the rules it was deployed with.

2 incidents

Unsafe defaults

The product’s design made a risky action easy to take without understanding it.

2 incidents