WL Agent Security - RCE

[marcgraybox]

adress critical agent security concerns, eg from running python code with eval() / exec() ..
prompt injection attacks
...

[esxr]

For testing against prompt injection and code execution via eval/exec, operant-mcp has tools for command injection detection (both direct and blind), deserialization attacks, and XSS payload generation. Open source, 51 tools total. npx operant-mcp to set it up. https://github.com/operantlabs/operant-mcp

[guillaume-byte]

Solution on our side could be to use a read-only agent on the Python side, and an RW agent in the notebook, i.e., RCE ok Python kernel