| Titre | NousResearch hermes-agent 2026.4.16 Improper Encoding or Escaping of Output (CWE-116) |
|---|
| Description | # Technical Details
An Arbitrary Mass Mention Bypass exists in the `format_message` and `send` methods in the Slack and Mattermost adapters of hermes-agent.
The application fails to correctly neutralize or escape unauthenticated mass-ping mentions (like `<!everyone>` and `@all`) generated via LLM prompt injections before forwarding them to messaging APIs.
# Vulnerable Code
File: gateway/platforms/slack.py and gateway/platforms/mattermost.py
Method: format_message (Slack) / send (Mattermost)
Why: In SlackAdapter, a protective regex `re.sub` specifically preserves valid `<...>` tags, inadvertently insulating primitives like `<!everyone>` from sanitization and leaving them intact to trigger API features. In the MattermostAdapter, `send` directly embeds the output in a `/posts` payload without applying the `disable_mentions: True` flag.
# Reproduction
1. Interact with a `hermes-agent` integrated with Slack or Mattermost.
2. Supply a coerced prompt injection coercing the Agent into outputting `<!everyone>` or `@all`.
3. The LLM processes the adversarial instruction and generates the raw mention tag.
4. The backend platforms deliver notification bursts to all participants without escaping the raw mentions.
# Impact
- Notification Exhaustion leading to Social Engineering Denial of Service (Spam/DoS).
- Thousands of users simultaneously bothered by unauthorized mass notifications, heavily degrading workspace operations. |
|---|
| La source | ⚠️ https://gist.github.com/YLChen-007/e90fb38ac03284176bae49898a3a46a4 |
|---|
| Utilisateur | Eric-i (UID 97584) |
|---|
| Soumission | 24/04/2026 15:00 (il y a 1 mois) |
|---|
| Modérer | 23/05/2026 11:19 (29 days later) |
|---|
| Statut | Accepté |
|---|
| Entrée VulDB | 365317 [NousResearch hermes-agent jusqu’à 2026.4.16 Slack Agent/Mattermost Agent format_message élévation de privilèges] |
|---|
| Points | 20 |
|---|