godmode
Audited by Socket on Sep 14, 2026
7 alerts found:
Malwarex2Securityx2Anomalyx3SUSPICIOUS/HIGH-RISK skill. The footprint is internally consistent with its stated purpose, but that purpose is to defeat LLM safeguards, persist jailbreaks in agent config, and optimize for less-filtered harmful responses. No clear evidence of credential theft or malware delivery, but the skill meaningfully increases misuse capability and should be treated as a high-risk offensive prompt-injection tool.
The code is an explicit LLM jailbreak automation tool, not a conventional host malware payload. It intentionally bypasses model safety controls, sends built-in or user-supplied prompts to a configurable remote API, and persists successful jailbreak settings. The import-time exec() of neighboring scripts is a high-risk code-execution pattern, and the unrestricted base_url creates a potential API-key disclosure risk if controlled by an attacker. The fragment contains no direct reverse shell, cryptominer, destructive payload, or clear system-data exfiltration. It should be treated as dangerous for applications requiring model safety and reviewed together with the executed helper scripts before use.
This module implements a high-risk dynamic code execution loader. It reads and execs Python source files from a directory derived from the HERMES_HOME environment variable, then exports selected functions/constants from those executed scripts into the current module namespace. Even without seeing the loaded scripts’ contents, the design creates a strong supply-chain/local-compromise pathway: if an attacker can influence HERMES_HOME or modify the target directory files, they can execute arbitrary code with the importing process’s privileges. The snippet contains no explicit networking/exfiltration/persistence logic itself, but it provides the necessary execution primitive to do so via the loaded scripts.
No evidence of traditional host-compromise malware in this module (no execution primitives, persistence, file manipulation, or local data theft). The primary security concern is behavioral and privacy-related: it forwards user-supplied prompts to a third-party LLM gateway (potentially many models concurrently), prints the selected output (risking sensitive data exposure in logs), and includes hardcoded jailbreak-style system prompts plus refusal/hedge heuristics that steer toward non-refusal/unfiltered content. Treat as an LLM proxy with policy-evasion intent and non-trivial data-sharing risk rather than as a stealth malware package.
This module is a dual-use text obfuscation engine focused on detecting cybersecurity/malware-related keywords and rewriting them with stealthy Unicode/zero-width and encoding transformations (leet, homoglyph-like mappings, ZWJ/ZWNJ insertion, fullwidth, Base64/hex, etc.). There is no direct malware behavior in this fragment (no exec, network, or file/process actions), but the design strongly supports evasion of keyword-based filters or security tooling. Treat as moderate-to-high supply-chain risk depending on how and where it is used.
The fragment is an openly documented collection of model-jailbreak prompts and agent-invocation instructions. It does not itself contain clear malware or data theft, but it is designed to bypass model safety controls and includes a risky dynamic exec() example. Treat it as untrusted red-team content; do not install or execute the referenced script without verifying its provenance and contents.
This fragment is a high-risk runtime loader: it reads a Python file from a filesystem location determined by the HERMES_HOME environment variable (defaulting to ~/.hermes) and executes the loaded contents via exec(...), with no integrity verification or safeguards. While the visible demo code only performs refusal detection/score printing, the exec primitive means the executed script could implement arbitrary behavior. Without reviewing godmode_race.py, exact malicious intent cannot be confirmed, but the security risk is substantial and warrants immediate scrutiny and containment.