At Broad Peak Cyber, we’re increasingly advising UK organisations on the risks associated with agentic AI, a new class of AI systems capable of taking real-world actions.
Frameworks like OpenClaw are accelerating innovation, but they are also introducing fundamental shifts in the cyber threat model—often faster than organisations can secure them.
Why Agentic AI Changes the Rules
Traditional AI systems generate outputs.
Agentic AI systems:
- Access files and systems
- Use APIs and credentials
- Execute commands
- Make decisions autonomously
This creates a new reality:
AI is no longer just analysing your environment; it is actively operating inside it.
When deployed insecurely, these agents can become:
- Entry points for attackers
- Vehicles for lateral movement
- Tools for executing malicious actions
What We’re Seeing in the Field
Recent issues with OpenClaw highlight broader risks we are now seeing across UK environments.
1. Internet-Exposed AI Agents
Many deployments have been:
- Exposed to the internet due to misconfiguration
- Bound to open interfaces instead of localhost
This creates direct access to:
- Agent capabilities
- Connected systems
- Stored credentials
In practical terms: If an attacker reaches the agent, they inherit its access.
2. Agent Hijacking via Trusted Channels
One of the most concerning patterns involves agent hijacking.
Attackers can:
- Exploit trust assumptions (e.g. localhost connections)
- Use malicious websites to interact with agents
- Take control without obvious user interaction
This allows:
- Command execution
- Data exfiltration
- Privileged actions under legitimate identity
3. Supply Chain and Shadow Deployment Risks
We’re also seeing:
- Unauthorised installations via compromised packages
- Developers deploying agents without security oversight
- No central visibility or control
For security teams, this creates a major challenge—you can’t secure what you can’t see.
Why Agentic AI Is a Threat Multiplier
These risks go beyond a single framework.
Agentic AI combines:
- Untrusted external inputs
- Third-party code
- Persistent memory
- High-privilege execution
In one loop.
An agent can:
- Read malicious content
- Interpret it as instructions
- Store it
- Execute actions later
All using legitimate access.
As highlighted by Microsoft security research: This model should be treated as untrusted code execution with credentials.
The Real Risk: Insecure Deployment
Across UK organisations, the biggest risks are not zero-day exploits.
They are:
- Over-privileged agents
- Poor isolation
- Internet exposure
- Weak governance
In other words: The problem is not the technology, it’s how it’s deployed.
How We Help UK Organisations Reduce Risk
At Broad Peak Cyber, our approach is simple:
Treat agentic AI as high-risk, untrusted execution by default.
1. Aggressive Isolation
We ensure agents run:
- In dedicated virtual machines or containers
- Separate from core business systems
- With no direct access to sensitive environments
2. Least Privilege by Design
We enforce:
- Narrowly scoped permissions
- Short-lived credentials
- No standing access to critical systems
If compromised, the blast radius is contained.
3. Zero Trust for AI
We apply Zero Trust principles:
- No implicit trust in inputs
- Continuous validation of actions
- Segmented access to systems
This is critical in preventing:
- Prompt injection attacks
- Malicious instruction execution
4. Monitoring, Logging and Rapid Recovery
Agentic AI requires a different mindset:
- Assume compromise is possible
- Monitor all agent actions
- Maintain full audit trails
We implement:
- Real-time monitoring
- Behavioural analytics
- Rebuild-on-compromise strategies
5. Governance and Visibility
Shadow AI is one of the biggest risks we see.
We help organisations:
- Discover where agents are running
- Maintain centralised inventory
- Enforce deployment standards
- Align with compliance requirements
6. Secure Input Handling
We ensure:
- Validation of all external inputs
- Filtering of web, email and API data
- Protection against prompt manipulation
7. User and Developer Awareness
Technology alone won’t solve this.
We support:
- Developer training on secure AI deployment
- User awareness around risks
- Clear policies on AI usage
What UK Organisations Should Do Now
If you’re exploring agentic AI, ask:
- Do we know where agents are deployed?
- Are they isolated from critical systems?
- Are privileges tightly controlled?
- Can we monitor and audit their actions?
- Do we have a plan if an agent is compromised?
If the answer to any of these is unclear, your risk exposure is already higher than you think.
Final Thought
Agentic AI represents a powerful leap forward, but also a significant expansion of the attack surface.
Frameworks like OpenClaw are not inherently insecure.
But without the right controls, they can:
- Amplify existing vulnerabilities
- Introduce new attack paths
- Blur the line between trusted and untrusted execution
At Broad Peak Cyber, we help UK organisations adopt AI securely and strategically so innovation can move forward without exposing the business to unnecessary risk.
Because in this new era, it’s not just about securing systems—it’s about securing the decisions machines are now making on your behalf.
