Executive brief
NousResearch hermes-agent is a tool used for managing AI agent skills and behaviors. A vulnerability in its security filtering component allows attackers to bypass safety checks by slightly modifying malicious commands, such as inserting extra words into a prompt. This could allow an attacker to install malicious skills that override the agent's core instructions, potentially leading to unauthorized actions or data manipulation without the user's knowledge.
Technical details
A prompt injection vulnerability exists in the Skills Guard Multi-Word Prompt Handler within `agent/skills_guard.py`. The root cause is the use of rigid regular expressions for the `THREAT_PATTERNS` argument, specifically for patterns like 'system prompt override'. Because these patterns do not account for flexible whitespace or intervening words, an attacker can bypass the filter by injecting extraneous words (e.g., 'system prompt temporary override'). This bypass allows the installation of malicious skills that can override system prompts and alter agent behavior. The issue is addressed in version 0.15.0 by implementing more flexible regex matching.
Affected products
- NousResearch hermes-agent < 0.15.0
Timeline
- 2026-04-24: disclosed: Public disclosure via GitHub Gist and PoC release
- 2026-05-24: advisory: NVD publication date
- 2026-05-26: advisory: GitHub Advisory published
- 2026-06-30: patched: GitHub Advisory updated with patch information