The short version

Cape Town’s GovernX has launched AgentShield, a security add‑on that watches AI assistants and blocks malicious commands before they hit a database. The tool can stop rogue actions in under a millisecond, protecting sensitive data from prompt‑injection attacks.

What happened

Cape Town‑based GovernX released Custos AgentShield, a module that sits inside its Custos data‑protection platform.

The software watches autonomous AI agents, programmes that act across applications and databases without a human approving each step, and checks the commands they issue.

If an agent tries to delete or alter database records, retrieve unauthorised information, or open a risky network connection, AgentShield blocks the request and can terminate the connection in less than a millisecond.

Zero telemetry leaves South African borders

Brandon Fortune

The background

Prompt injection is a cyberattack where hidden instructions in documents or messages trick an AI assistant to act outside its intended task.

GovernX illustrates the risk with a scenario where an assistant summarising a support ticket is fed malicious instructions that make it try to pull customer records.

AgentShield is built to recognise the unauthorised change in behaviour and stop the database connection before any data is exposed.

What it means for owners and operators

Owners of data‑rich businesses can reduce the risk that a misbehaving AI agent harms their core systems.

Operators gain a tool that automatically blocks rogue commands, meaning they spend less time monitoring AI outputs manually.

The sub‑millisecond response time limits any potential damage, making the solution attractive for finance, telecoms and critical‑infrastructure firms that cannot afford downtime.

Because the product can run on‑premises or in air‑gapped environments, it avoids sending any telemetry outside South Africa, helping organisations meet strict data‑sovereignty rules.

The numbers

GovernX says AgentShield can interrupt a suspicious connection in less than one millisecond.

The network termination itself can happen in as little as 126 microseconds.

The broader Custos platform’s ransomware module, ShannonEntropy, responds in under 500 milliseconds.

What to do in your business tomorrow

  1. Review any AI assistants that interact with internal databases and map the commands they can issue.
  2. Patch your AI workflow with a security layer that validates each command against approved functions.
  3. Test the latency of your existing safeguards; if response times exceed a millisecond, consider faster solutions.

Related coverage

Sources