- Category
- AI red teaming
- Website
- mindgard.ai
- Last verified
- November 1, 2026
What is Mindgard?
Mindgard is an AI security company that describes itself as an offensive security platform for AI systems and agents. It says it secures AI models, agents and applications, and that it was spun out of more than a decade of AI security research at Lancaster University. Its platform follows an attacker's workflow: map the AI assets, attack them, help fix what is found and keep testing over time.
Key capabilities
- Recon: maps AI assets, models, agents, tools and behaviors the way an attacker would.
- Attack: adaptive attacks that find and validate exploitable vulnerabilities.
- Remediate and Assure: evidence, fix guidance, retesting and continuous assessment.
- Agent testing, including tool-using agents and systems connected through MCP and A2A.
- Runtime protection and red teaming services, with API, CI/CD and Burp Suite integrations.
Why it matters for humanoid robot security
Editorial analysis: Mindgard does not market itself for robots, but its focus on tool-using agents is directly relevant. In many humanoid designs, an LLM or agent layer plans tasks and calls tools that move the body. Research covered on this site has shown that jailbreaks and prompt injection can push such robots into unsafe actions. Automated red teaming of that agent layer, before it is connected to actuators, is one of the cheapest places to catch those failures.
Where it fits in a humanoid robot security program
Analysis: Mindgard fits at the AI model and agent layer of a humanoid, the part that interprets instructions and decides which skills or tools to call. It is most useful during design and testing, by running attack campaigns against the robot's planner or assistant in a simulator or staging setup, and during operations as a recurring test after model updates. It does not cover firmware, controllers or the site network.
Who it is for
- Robot makers building LLM or agent-based task planners
- AI teams that must show red-team evidence before a release
- Security teams testing third-party AI agents connected to physical systems
Questions to ask Mindgard
- Can you red team an agent whose tools trigger physical actions, using a simulator or mock tools?
- Do your attacks cover multimodal inputs such as images and audio, not just text?
- How do findings map to OWASP and MITRE ATLAS categories?
- Can tests run automatically in CI/CD after each model or prompt change?
Similar listings
- HiddenLayer: AI and model security
- Promptfoo: AI red teaming and evaluation
- F5 AI Guardrails (CalypsoAI): AI runtime security
Related on Humanoid Threats
- Can LLM-controlled robots be jailbroken?
- Prompt injection against robot VLA models
- Browse the full Humanoid Robot Security Directory
Sources
This listing is editorial and free; inclusion is not an endorsement. Work at Mindgard? Claim or update this listing.
Get the weekly Humanoid Threats Brief → Subscribe


