The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a user-defined 'test_command' to autonomously verify code fixes. This allows the agent to execute arbitrary shell commands on the system as part of its internal logic.
[PROMPT_INJECTION]: The skill contains strong directives to 'LOOP FOREVER,' 'NEVER stop,' and 'NEVER ask permission to continue.' These are explicit attempts to override standard interaction safety protocols and bypass user oversight.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the processing of untrusted data. * Ingestion points: The agent reads 'run.log' and 'results.tsv' which contain raw output and crash logs from functions processing adversarial inputs. * Boundary markers: There are no explicit instructions to the agent to ignore or delimit instructions found within the test outputs. * Capability inventory: The skill has permissions to write to the file system (modifying source code and adding tests) and execute shell commands. * Sanitization: The skill recommends using 'grep' to parse logs rather than reading them directly, which provides limited mitigation against malicious content embedded in large log files.

fuzz-fixing