ask-peer
Fail
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: HIGHPRIVILEGE_ESCALATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [PRIVILEGE_ESCALATION]: The skill instructs the agent to intentionally bypass safety controls and user-approval mechanisms.
- Evidence in
references/ask-claude.md: Instructs the use of--permission-mode bypassPermissionsfor write tasks, which disables the platform's requirement for user confirmation before executing actions. - Evidence in
references/ask-codex.md: Instructs the use of-s danger-full-accesscombined with-c 'approval_policy="never"', which grants the peer agent unrestricted access without requiring user review for its operations. - [COMMAND_EXECUTION]: The skill relies on the execution of shell commands through the
claudeandcodexCLI tools. - Evidence in
references/ask-claude.mdandreferences/ask-codex.md: Commands use subprocess-style execution to pass the workspace context, model configuration, and user-supplied prompts to the external CLIs. - [INDIRECT_PROMPT_INJECTION]: The skill defines a high-risk attack surface by forwarding unvalidated user input to peer agents running with the aforementioned elevated privileges.
- Ingestion points: The
argument-hintinSKILL.mdcaptures user-provided questions or tasks. - Boundary markers: The bash templates in the reference files do not implement prompt delimiters or instructions to ignore embedded commands within the user input.
- Capability inventory: The peer agents are granted
Read, Grep, Globaccess by default and full write access (bypassPermissions/danger-full-access) when implementation is requested. - Sanitization: No sanitization or validation is performed on the prompt content (
$P) before it is piped to the peer CLIs. - [DATA_EXFILTRATION]: The skill facilitates the transfer of local project data to external model providers.
- Evidence in
SKILL.md: The 'Shared contract' mandates including the workspace root and relevant repository instructions in the external request. - Evidence in
SKILL.md: The agent is instructed to read local configuration files, specifically${CODEX_HOME:-$HOME/.codex}/config.toml, to extract model identifiers.
Recommendations
- AI detected serious security threats
Audit Metadata