I developed a tool that can catch AI cheating. Is it viable for research?
Like many people, I’m developing a harness for agentic AI to combat context bloat and enable autonomous code execution. It consists of three layers:
- the core harness with hooks, scripts, and skills - a zero dependency ticket system similar to Jira - a SQL database that combines the harness and the ticket system
The database gave me an audit layer to actually check what the agent did. Through this I found cheating, which is a known issue but I can’t find anything open source outside regulatory or academic research that has this level of granularity. For example, I have a PreToolUse with a Deny if the code for commands like rm -rf or sudo. The AI got around the hook by writing char(114)+char(119) -rf, and su””do. The compiler knew what to do, outside the hook's reach so it went through.
It’s not reward hacking, it’s closer to spec gaming or illusory compliance. I’m not a researcher and I’m not pretending to be, but I have to wonder if the tool is valuable to the research community.
I would love to hear feedback, or just tell me I found what is already developed so I don’t chase something that no one needs.
I developed a tool that can catch AI cheating. Is it viable for research?
Like many people, I’m developing a harness for agentic AI to combat context bloat and enable autonomous code execution. It consists of three layers:
- the core harness with hooks, scripts, and skills
- a zero dependency ticket system similar to Jira
- a SQL database that combines the harness and the ticket system
The database gave me an audit layer to actually check what the agent did. Through this I found cheating, which is a known issue but I can’t find anything open source outside regulatory or academic research that has this level of granularity. For example, I have a PreToolUse with a Deny if the code for commands like rm -rf or sudo. The AI got around the hook by writing char(114)+char(119) -rf, and su””do. The compiler knew what to do, outside the hook's reach so it went through.
It’s not reward hacking, it’s closer to spec gaming or illusory compliance. I’m not a researcher and I’m not pretending to be, but I have to wonder if the tool is valuable to the research community.
I would love to hear feedback, or just tell me I found what is already developed so I don’t chase something that no one needs.