handoff
Warn
Audited by Gen Agent Trust Hub on Nov 19, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill retrieves configuration data from
docs/agents/workspace.mdanddocs/agents/handoffs.mdlocated within the active repository. These files are untrusted ingestion points as they are part of the repository content. Malicious data or instructions placed in these files can influence the agent's behavior.\n- [COMMAND_EXECUTION]: Instructions inSKILL.mddirect the agent to execute shell commands specified in the repository's settings, such as theWorkspace status command. This creates a vulnerability where an attacker can achieve arbitrary command execution by defining malicious commands in the markdown-based configuration files.\n- [DATA_EXFILTRATION]: By design, the skill aggregates repository state, branch information, and 'machine-local knowledge' into a handoff document and moves it to a cloud-synced directory. While the skill includes prompts to avoid secrets, the automated collection and external transmission of local session metadata represents a data exfiltration vector.\n- [DYNAMIC_EXECUTION]: The skill utilizes a settings framework (defined insettings/README.md) that allows defining values as executable commands wrapped in backticks. The agent is instructed to 'complete the settings' by reading these values and subsequently executing them, which facilitates the execution of dynamically defined code from the repository environment.
Audit Metadata