floating-terminal
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill's primary function is to execute arbitrary shell commands via the Orca IDE CLI (
orca terminal create). It specifically targets high-privilege or interactive commands that require a TTY or human presence, such assudoor biometric authentication. - [DYNAMIC_EXECUTION]: The
scripts/orca-floatutility dynamically generates two shell scripts (launch.shandcmd.sh) in a temporary directory (/tmp/orca-float.XXXXXX) created with restricted permissions (mode 700). The script correctly usespython3'sshlex.quoteto sanitize user-provided arguments before they are interpolated into the generated shell scripts, effectively mitigating common command injection vectors during the generation process. - [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes external data by reading the terminal output from commands executed in the floating terminal. This output is returned to the agent's context, which could be exploited to deliver malicious instructions if the command interacts with untrusted files or remote services.
- Ingestion points:
scripts/orca-floatreads the log of the terminal session ($rundir/log) to return the final lines of output. - Boundary markers: The terminal output is encapsulated as a string value within a structured JSON object returned by the tool.
- Capability inventory: The skill possesses full shell execution capabilities within the user's terminal environment.
- Sanitization: The script performs regex-based cleaning of ANSI escape codes and terminal artifacts but does not sanitize the semantic content of the terminal output for potential prompt injection patterns.
- [PROMPT_INJECTION]: The
SKILL.mdfile contains a 'Self-Evolving Skill' block that instructs the agent to 'fix this file and the script immediately' if discrepancies are found. This meta-instruction encourages the agent to modify its own code, which could be manipulated by an attacker to induce the agent to introduce malicious logic under the guise of a 'fix'.
Audit Metadata