>
AgentArmor
Hardening LLM agents against prompt injection.
A control checklist that moves every security decision outside the model, where an attacker's text cannot argue with it.
$
git clone https://github.com/MuhammadMurtuzaHussain/agent-armor ~/.claude/skills/agent-armorMade at Dublin AI WeekMaps to the OWASP LLM Top 10, NIS2 and ISO 27001.
The pair, live
Watch an attempt get blocked.
The same seven attack classes, one per line. On AgentArmor the control stops them. Hover a line to scan the redacted payload.
attack-vs-defense
ATTEMPT -> BLOCKEDTool-description / MCP poisoning
> tool.description <= hidden parameter instruction
// scanning vector 01 / 07
BLOCKEDschema caps the field to an enum
hover to scan
Seven controls, one per attack class
Any guardrail the model enforces by reading untrusted text is bypassable. Each control moves the decision to a layer an attacker cannot argue with.
vector 01 / 07
Constrain tool metadata at the schema
Treat tool metadata as untrusted data. Constrain at the schema level what each argument can carry, so a free-text field cannot hold PII.
Install
Clone it into your Claude skills directory, then restart the session. The clone ships the full SKILL.md.
bash
$ git clone https://github.com/MuhammadMurtuzaHussain/agent-armor ~/.claude/skills/agent-armor