Researchers at Varonis disclosed CoSnitch, a flaw that made Microsoft Copilot exfiltrate data without tripping obvious alarms — reportedly the third such Copilot flaw this year. Any Cyprus business connecting AI assistants to its data should take note.
Our analysis
Three exfiltration flaws in one product, in one year, is a pattern, not bad luck. AI assistants wired into corporate data are a recurring, high-value target, and the stealth is the scary part: this isn’t the noisy data theft your DLP was built to catch. For Cyprus firms handling EU customer data, a silent AI-driven leak is still a GDPR incident.
What you should do
- Treat any AI assistant with data access as manipulable by untrusted content.
- Scope Copilot and agent permissions to the minimum data and actions needed.
- Monitor agent tool calls and outbound data flows, not just user prompts.
- Red-team AI deployments before launch and after each major change.
How CyprusPentest helps: our AI agent penetration testing and prompt injection testing probe these exfiltration paths.
Source: eSecurity Planet.
How the exfiltration actually works
The attack needs no malware. An attacker plants instructions inside content the assistant will later read — a shared document or email. When a legitimate user asks a normal question, the assistant retrieves that poisoned content and quietly follows the hidden instructions, sending sensitive data out through a channel it is allowed to use.
This is indirect prompt injection, the hardest LLM risk to defend. The defence is to treat every ingested document as untrusted and to keep authorisation and data-access decisions in your own code, never in the model.
What this means for Cyprus businesses
An AI assistant that can be tricked into leaking data needs no malware — just crafted content it reads. For Cyprus businesses adopting Copilot-style tools, any assistant touching personal or financial data is now an exfiltration surface, and a leak is a GDPR matter.
How to reduce your exposure
- Treat every AI assistant with data access as untrusted and scope its permissions tightly.
- Apply least-privilege to what it can read and where its output can go.
- Monitor and log assistant actions and data access.
- Test against indirect prompt injection before rollout.
Assess your AI tools with prompt injection testing and AI agent penetration testing — get a fixed-price quote for your Cyprus business.