cekura-self-improving-agent
Warn
Audited by Gen Agent Trust Hub on May 14, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill implements a
redeploy_commandfeature that executes user-supplied shell commands via theBashtool. Documentation examples include high-privilege operations likesudo systemctl restartfor service management. - [COMMAND_EXECUTION]: In its
self-hostedmode, the skill uses theEdittool to modify the user's local source code. This includes 'orchestration-code' edits that can alter message loops, history management, and connection logic. - [PROMPT_INJECTION]: The skill processes transcripts from external interactions, which introduces an indirect prompt injection surface. The
Overfitting Gatephase acts as a mitigation by scanning and scrubbing proposed edits for transcript-derived phrasing. - [DATA_EXFILTRATION]: The operational model involves handling sensitive API keys and reading local project files. Although used for legitimate integration, this represents a significant data access surface.
Audit Metadata