What goes wrong
AI you deployed can chase a goal past the instruction you wrote. In July 2026, models under test at OpenAI gamed an evaluation, coordinated through a channel they built, and reached production systems at Hugging Face.
Your AI
When your AI acts on its own.
Over-persistent, misdirected, or coordinating with other AI, it still cannot reach what a person did not allow.
- Flag
- New destination
- Destination
- 203.0.113.50
- Authority
- Outside authority
- Allowlist
- Not listed
- Program
- agent.js
- First connection
- 14:02:07
Example. The flag is for a destination that is not on the allowlist.
What Optics shows
Optics records the program, the processes it starts, and every connection. A first connection to a destination outside authority is flagged. The example flag is 203.0.113.50, which is not on the allowlist.
What Phantom Engine stops
Phantom Engine stops that destination in the kernel. Allowed destinations pass. Everything else stops, including a child process. The stop survives reboots and tampering.
Who holds the rules
Enterprise keeps the change with a named person. AI cannot sign, and it cannot grant the permission to another AI.