SecurityQuiet 84d · day 86
Varonis researchers trick Copilot into disclosing its own security flaws
Social engineering attack exposes undocumented parameters that allow malicious prompts to auto-execute without user consent.
What to know
- Researchers used 'meta-hacking'—social engineering the AI's reasoning engine—to trick Copilot into disclosing its own security mechanisms.
- The attack exploits an undocumented autorun parameter that causes malicious prompts to execute automatically in a victim's authenticated session with no user confirmation.
- Copilot can be tricked via phishing URLs to send sensitive data to external servers and poison its persistent memory.
- Microsoft was notified in December 2025 and planned to patch the vulnerability and assign a CVE designation.
“Critically, Copilot also described its own protections against this behavior and explained why the parameter 'no longer works.' Those explanations were precise and technically detailed. When we tested the parameter exactly as Copilot described it, the one it told us was disabled, executed.”
Varonis Threat Labs · The Register
Varonis Threat Labs Security researchersMicrosoft Copilot developer
The record 1 articles and posts · last 30 days