codex-tools-account-manager
Fail
Audited by Gen Agent Trust Hub on Sep 16, 2026
Risk Level: HIGHPRIVILEGE_ESCALATIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [PRIVILEGE_ESCALATION]: The skill instructs users to run
sudo spctl --master-disableon macOS. This command globally disables Gatekeeper, which is a critical security control designed to prevent the execution of unverified or malicious software. Disabling it significantly compromises the system's security integrity.\n- [EXTERNAL_DOWNLOADS]: The instructions direct users to download and install executable files from a repository hosted by a third-party account (170-carry/codex-tools) that is not associated with a verified or well-known organization.\n- [COMMAND_EXECUTION]: The skill recommends usingsudo xattr -r -d com.apple.quarantineto bypass security flags on downloaded binaries, encouraging users to ignore operating system security warnings.\n- [DATA_EXFILTRATION]: The tool features built-in integration withcloudflaredto create public tunnels to a local proxy server. This proxy server manages sensitive session data, includingaccess_tokenandrefresh_tokenfor Codex accounts. Exposing a local credential-handling service to the public internet via a tunnel significantly increases the risk of data interception and account compromise.\n- [INDIRECT_PROMPT_INJECTION]: The skill sets up an environment where an agent interacts with untrusted external data via an API proxy.\n - Ingestion points: The skill processes data through an OpenAI-compatible
/v1endpoint (SKILL.md).\n - Boundary markers: None identified in the provided instructions to differentiate between data and instructions.\n
- Capability inventory: The skill has the ability to terminate and launch editor applications, modify local configuration files, and initiate network tunnels (SKILL.md).\n
- Sanitization: There are no instructions for validating or sanitizing the content passing through the proxy endpoint.\n- [METADATA_POISONING]: There is a mismatch between the author listed in the metadata (
ara.so) and the GitHub repository owner (170-carry). Furthermore, the context provided identifies the author as 'reason-machines', contradicting the skill's own claims, which may lead to confusion regarding the software's origin and safety.
Recommendations
- AI detected serious security threats
Audit Metadata